Dataset Viewer
Auto-converted to Parquet Duplicate
Model
stringclasses
5 values
Status
stringclasses
2 values
Task
stringclasses
6 values
Episode
int64
0
899
Prompt
stringclasses
356 values
Scene Setup
stringclasses
363 values
Expected Target / Action
stringclasses
143 values
Observations
stringlengths
8
227
Condition
stringclasses
17 values
Scene Family
stringclasses
440 values
Comparison / Subtype
stringclasses
116 values
Reference Episode
float64
0
168
Original Training Episode(s)
stringclasses
346 values
Manifest Alignment
stringclasses
62 values
Collection Block
stringclasses
12 values
Setup Change from Previous
stringclasses
97 values
Design Note
stringclasses
272 values
Reference Media Filename
float64
Object Count
float64
2
5
Valid Candidate Count
float64
0
1
Bowl Start
stringclasses
1 value
Queried State
stringclasses
4 values
Queried Category
stringclasses
5 values
Contrast Pair
stringclasses
116 values
Initial Bowl Count
float64
0
4
Goal Count
float64
0
3
Action Direction
stringclasses
3 values
Target Position
stringclasses
4 values
Query Concept
stringclasses
20 values
Relation
stringclasses
4 values
Target Object
stringclasses
9 values
Reference Object
stringclasses
5 values
Juggling Ball Role
stringclasses
2 values
Candidate Category
stringclasses
5 values
Target Side
stringclasses
3 values
Separation (cm)
float64
8
16
Tested Cell / Role
stringclasses
16 values
Source Workbook
stringclasses
6 values
Source Result Sheet
stringclasses
11 values
Source Result Row
int64
2
171
Source Plan Row
int64
2
171
VLA-0 (Epoch 12)
Failure
Counting
0
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [blue pen, yellow block, green pencil sharpener]
ADD any one table object; finish with 1 object in the bowl
didn't move
In-distribution concept baseline
ID-F01
In-distribution: one via trained 0→1 addition
null
801
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 801.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
2
2
VLA-0 (Epoch 12)
Failure
Counting
1
Adjust the bowl so that it contains two objects.
Bowl objects: [yellow block] | Table objects: [blue pen, green pencil sharpener]
ADD any one table object; finish with 2 objects in the bowl
tried to pick the blue pen, but didn't even open the gripper
In-distribution concept baseline
ID-F01
In-distribution: two via trained 1→2 addition
null
809
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 809.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
3
3
VLA-0 (Epoch 12)
Failure
Counting
2
Leave exactly three objects in the bowl.
Bowl objects: [blue pen, yellow block, green pencil sharpener, airplane toy without propellers] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
started gripping in middle of workspace like it's going for a soda can or cup
In-distribution concept baseline
ID-F01
In-distribution: three via trained 4→3 removal
null
843
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 843.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
4
4
VLA-0 (Epoch 12)
Failure
Counting
3
Adjust the bowl so that it contains zero objects.
Bowl objects: [blue pen] | Table objects: [green pencil sharpener, airplane toy without propellers]
REMOVE the only bowl object; finish with 0 objects in the bowl
didn't move
In-distribution concept baseline
ID-F01
In-distribution: zero via trained 1→0 removal
null
826
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 826.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
5
5
VLA-0 (Epoch 12)
Failure
Counting
4
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [yellow block, blue pen]
ADD any one table object; finish with 1 object in the bowl
went to middle of workspace and couldn't grab anything
In-distribution concept baseline
ID-F02
In-distribution: one via trained 0→1 addition
null
852
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 852.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
6
6
VLA-0 (Epoch 12)
Failure
Counting
5
Adjust the bowl so that it contains two objects.
Bowl objects: [green pencil sharpener] | Table objects: [yellow block, airplane toy without propellers]
ADD any one table object; finish with 2 objects in the bowl
went to the yellow block, but couldn't grab it
In-distribution concept baseline
ID-F02
In-distribution: two via trained 1→2 addition
null
859
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 859.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
7
7
VLA-0 (Epoch 12)
Failure
Counting
6
Leave exactly three objects in the bowl.
Bowl objects: [yellow block, green pencil sharpener, airplane toy without propellers, blue pen] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
Went to bowl but couldn't grab anything
In-distribution concept baseline
ID-F02
In-distribution: three via trained 4→3 removal
null
893
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 893.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
8
8
VLA-0 (Epoch 12)
Failure
Counting
7
Adjust the bowl so that it contains zero objects.
Bowl objects: [yellow block] | Table objects: [blue pen]
REMOVE the only bowl object; finish with 0 objects in the bowl
didn't move
In-distribution concept baseline
ID-F02
In-distribution: zero via trained 1→0 removal
null
878
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 878.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
9
9
VLA-0 (Epoch 12)
Failure
Counting
8
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [green pencil sharpener, yellow block]
ADD any one table object; finish with 1 object in the bowl
went towards the yellow block which is on the right side of workspace but didn't grab it
In-distribution concept baseline
ID-F03
In-distribution: one via trained 0→1 addition
null
902
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 902.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F03
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
10
10
VLA-0 (Epoch 12)
Failure
Counting
9
Adjust the bowl so that it contains two objects.
Bowl objects: [blue pen] | Table objects: [green pencil sharpener, airplane toy without propellers]
ADD any one table object; finish with 2 objects in the bowl
went to airplane toy but couldn't grab it
In-distribution concept baseline
ID-F03
In-distribution: two via trained 1→2 addition
null
909
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 909.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F03
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
VLA0 Results
11
11
VLA-0 (Epoch 12)
Failure
Ordinal Position
10
Put the second object from the right in the bowl.
L1: blue tape ball, L2: airplane toy with propellers
blue tape ball
just jittered around the home position
In-distribution concept baseline
ID-EXACT01
Exact training recreation — source episode 658
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 658; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 658: blue tape ball | airplane toy with propellers.
null
2
null
null
null
null
None — independent exact training recreation
null
null
null
L1
second from right
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
2
2
VLA-0 (Epoch 12)
Failure
Ordinal Position
11
Put the second object from the left in the bowl.
L1: blue tape ball, L2: airplane toy with propellers
airplane toy with propellers
went to airplane toy but didn't pick it up
In-distribution concept baseline
ID-EXACT02
Exact training recreation — source episode 619
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 619; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 619: blue tape ball | airplane toy with propellers.
null
2
null
null
null
null
None — independent exact training recreation
null
null
null
L2
second from left
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
3
3
VLA-0 (Epoch 12)
Failure
Ordinal Position
12
Move the second item counted from the left into the bowl.
L1: airplane toy with propellers, L2: green highlighter
green highlighter
went to airplane toy but didn't pick it up
In-distribution concept baseline
ID-EXACT03
Exact training recreation — source episode 719
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 719; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 719: airplane toy with propellers | green highlighter.
null
2
null
null
null
null
None — independent exact training recreation
null
null
null
L2
second from left
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
4
4
VLA-0 (Epoch 12)
Failure
Ordinal Position
13
Move the second item counted from the right into the bowl.
L1: yellow pen, L2: airplane toy with propellers
yellow pen
moved to middle of the workspace again
In-distribution concept baseline
ID-EXACT04
Exact training recreation — source episode 758
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 758; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 758: yellow pen | airplane toy with propellers.
null
2
null
null
null
null
None — independent exact training recreation
null
null
null
L1
second from right
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
5
5
VLA-0 (Epoch 12)
Failure
Ordinal Position
14
Put the third object from the right in the bowl.
L1: airplane toy with propellers, L2: yellow pen, L3: blue tape ball
airplane toy with propellers
went for yellow pen but didn't grasp it. seems to just be moving the arm to the middle of the workspace
In-distribution concept baseline
ID-EXACT05
Exact training recreation — source episode 604
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 604; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 604: airplane toy with propellers | yellow pen | blue tape ball.
null
3
null
null
null
null
None — independent exact training recreation
null
null
null
L1
third from right
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
6
6
VLA-0 (Epoch 12)
Failure
Ordinal Position
15
Put the object that is first from the right in the bowl.
L1: airplane toy with propellers, L2: green highlighter, L3: blue tape ball
blue tape ball
went for blue tape ball which actually wasn't in center of workspace; the green pen was in the center. couldn't pick up the tape ball
In-distribution concept baseline
ID-EXACT06
Exact training recreation — source episode 679
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 679; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 679: airplane toy with propellers | green highlighter | blue tape ball.
null
3
null
null
null
null
None — independent exact training recreation
null
null
null
L3
first from right
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
7
7
VLA-0 (Epoch 12)
Failure
Ordinal Position
16
Put the object that is third from the left in the bowl.
L1: green highlighter, L2: blue tape ball, L3: yellow pen
yellow pen
went to green highlighter but didn't pick it up
In-distribution concept baseline
ID-EXACT07
Exact training recreation — source episode 784
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 784; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 784: green highlighter | blue tape ball | yellow pen.
null
3
null
null
null
null
None — independent exact training recreation
null
null
null
L3
third from left
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
8
8
VLA-0 (Epoch 12)
Failure
Ordinal Position
17
Put the first object from the left in the bowl.
L1: green highlighter, L2: airplane toy with propellers, L3: blue tape ball
green highlighter
went towards the green highlighter, but couldn't pick it up
In-distribution concept baseline
ID-EXACT08
Exact training recreation — source episode 706
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 706; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 706: green highlighter | airplane toy with propellers | blue tape ball.
null
3
null
null
null
null
None — independent exact training recreation
null
null
null
L1
first from left
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
9
9
VLA-0 (Epoch 12)
Failure
Ordinal Position
18
Put the fourth object from the right in the bowl.
L1: blue tape ball, L2: green highlighter, L3: airplane toy with propellers, L4: yellow pen
blue tape ball
went for green highlighter/sort of towards the blue tape ball
In-distribution concept baseline
ID-EXACT09
Exact training recreation — source episode 652
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 652; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 652: blue tape ball | green highlighter | airplane toy with propellers | yellow pen.
null
4
null
null
null
null
None — independent exact training recreation
null
null
null
L1
fourth from right
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
10
10
VLA-0 (Epoch 12)
Failure
Ordinal Position
19
Put the second object from the left in the bowl.
L1: blue tape ball, L2: airplane toy with propellers, L3: yellow pen, L4: green highlighter
airplane toy with propellers
went towards airplane toy but didn't grasp it
In-distribution concept baseline
ID-EXACT10
Exact training recreation — source episode 614
null
null
null
1 — In-distribution baselines
Independently recreate the prompt and ordered object list from training episode 614; do not reuse the preceding baseline scene.
Exact final prompt and complete left-to-right object order from explicit training episode 614: blue tape ball | airplane toy with propellers | yellow pen | green highlighter.
null
4
null
null
null
null
None — independent exact training recreation
null
null
null
L2
second from left
null
null
null
null
null
null
null
null
Eval_Ordinal_Reference_Ordering.xlsm
VLA0
11
11
VLA-0 (Epoch 12)
Failure
Physical State
20
Put the open white cup in the bowl.
white cup A (O, RU); white cup B (C, RU); glass spice container A (O, RU); yellow-white striped cup B (O, RU)
white cup A
went for striped cup and knocked it over
In-distribution concept baseline
SR-ID01
Exact training-supported state-object baseline
null
2
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated. All objects are upright, matching the context of every explicit open/closed training demonstration.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The orthogonal orientation axis is fixed: every object is RU.
null
4
1
Empty
O
white cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
2
2
VLA-0 (Epoch 12)
Failure
Physical State
21
Put the upright white cup in the bowl.
white cup A (O, RU); white cup B (O, UD)
white cup A
went for upside-down white cup and didn't even grip it
In-distribution concept baseline
SR-ID02
Exact training-supported state-object baseline
null
26
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The same-category candidates share the orthogonal C/O state (O).
null
2
1
Empty
RU
white cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
3
3
VLA-0 (Epoch 12)
Failure
Physical State
22
Move the closed yellow-white striped cup into the bowl.
yellow-white striped cup B (C, RU); yellow-white striped cup A (O, RU); white cup A (C, RU)
yellow-white striped cup B
correctly went for the closed striped cup but then drift to grabbing empty space between the striped cup and white cup. Also started moving to the left of the workspace, pushing the white cup.
In-distribution concept baseline
SR-ID03
Exact training-supported state-object baseline
null
51
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated. All objects are upright, matching the context of every explicit open/closed training demonstration.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The orthogonal orientation axis is fixed: every object is RU.
null
3
1
Empty
C
yellow-white striped cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
4
4
VLA-0 (Epoch 12)
Failure
Physical State
23
Move the upside-down yellow-white striped cup into the bowl.
yellow-white striped cup B (O, UD); yellow-white striped cup A (O, RU); brown cup B (C, RU); glass spice container A (O, RU)
yellow-white striped cup B
went for the brown cup and then kind of was moving in between the brown cup and upright striped cup
In-distribution concept baseline
SR-ID04
Exact training-supported state-object baseline
null
78
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The same-category candidates share the orthogonal C/O state (O).
null
4
1
Empty
UD
yellow-white striped cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
5
5
VLA-0 (Epoch 12)
Failure
Physical State
24
Put the open brown cup in the bowl.
brown cup B (O, RU); brown cup A (C, RU); glass spice container A (O, RU)
brown cup B
went for open glass container. had it between grippers, but didn't close the grippers
In-distribution concept baseline
SR-ID05
Exact training-supported state-object baseline
null
104
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated. All objects are upright, matching the context of every explicit open/closed training demonstration.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The orthogonal orientation axis is fixed: every object is RU.
null
3
1
Empty
O
brown cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
6
6
VLA-0 (Epoch 12)
Failure
Physical State
25
Put the upside-down brown cup in the bowl.
brown cup A (O, UD); brown cup B (O, RU)
brown cup A
null
In-distribution concept baseline
SR-ID06
Exact training-supported state-object baseline
null
125
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The same-category candidates share the orthogonal C/O state (O).
null
2
1
Empty
UD
brown cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
7
7
VLA-0 (Epoch 12)
Failure
Physical State
26
Put the upright brown cup in the bowl.
brown cup B (C, RU); brown cup A (C, UD)
brown cup B
grasped thin air
In-distribution concept baseline
SR-ID07
Exact training-supported state-object baseline
null
134
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The same-category candidates share the orthogonal C/O state (C).
null
2
1
Empty
RU
brown cup
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
8
8
VLA-0 (Epoch 12)
Failure
Physical State
27
Place the closed glass spice container in the bowl.
glass spice container A (C, RU); glass spice container B (O, RU); yellow-white striped cup A (C, RU)
glass spice container A
was facing closed spice container, but didn't do anything; kind of froze
In-distribution concept baseline
SR-ID08
Exact training-supported state-object baseline
null
151
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated. All objects are upright, matching the context of every explicit open/closed training demonstration.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The orthogonal orientation axis is fixed: every object is RU.
null
3
1
Empty
C
glass spice container
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
9
9
VLA-0 (Epoch 12)
Failure
Physical State
28
Move the open glass spice container into the bowl.
glass spice container B (O, RU); glass spice container A (C, RU); white cup B (O, RU)
glass spice container B
didn't move
In-distribution concept baseline
SR-ID09
Exact training-supported state-object baseline
null
156
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated. All objects are upright, matching the context of every explicit open/closed training demonstration.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The orthogonal orientation axis is fixed: every object is RU.
null
3
1
Empty
O
glass spice container
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
10
10
VLA-0 (Epoch 12)
Failure
Physical State
29
Put the upside-down glass spice container in the bowl.
glass spice container B (C, UD); glass spice container A (C, RU)
glass spice container B
went for upright glass spice container, but didn't grip it; was reaching around the right side of the container
In-distribution concept baseline
SR-ID10
Exact training-supported state-object baseline
null
176
Exact manifest episode: trained state-object cell, prompt, same-category candidate structure, and object membership match; physical layout is manually recreated.
1
Recreate the listed manifest scene with the bowl empty.
The queried state-object combination and same-category discrimination structure are explicitly represented in training. The same-category candidates share the orthogonal C/O state (C).
null
2
1
Empty
UD
glass spice container
null
null
null
null
null
null
null
null
null
null
null
null
null
null
Eval_State_Recognition.xlsm
VLA0 Results
11
11
VLA-0 (Epoch 12)
Failure
Referential Description
30
Move the crushed soda can lying between the green pencil sharpener and the green pen into the bowl.
crushed Coke can; crushed Pepsi can; green pencil sharpener; green pen
crushed Coke can
moved wrist up a bit but didn't move to a can
In-distribution concept baseline
RD-ID01
Exact training-supported relation × candidate-category baseline
null
430
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 4 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
4
1
Empty
null
null
null
null
null
null
null
between
null
null
null
null
crushed soda can
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
2
2
VLA-0 (Epoch 12)
Failure
Referential Description
31
Pick up the crushed soda can positioned between the airplane toy with propellers and the green pen, then place it in the bowl.
crushed Coke can; crushed Pepsi can; airplane toy with propellers; green pen; green pencil sharpener
crushed Coke can
went to coke can, but didn't open gripper to actually grab it; stayed pretty still when it was by the coke can
In-distribution concept baseline
RD-ID02
Exact training-supported relation × candidate-category baseline
null
432
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 5 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
5
1
Empty
null
null
null
null
null
null
null
between
null
null
null
null
crushed soda can
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
3
3
VLA-0 (Epoch 12)
Failure
Referential Description
32
Pick up the crushed soda can positioned between the airplane toy with propellers and the green pencil sharpener, then place it in the bowl.
crushed Coke can; crushed Pepsi can; airplane toy with propellers; green pencil sharpener
crushed Coke can
didn't start moving to the coke can until end of episode; once got gripper around coke can, didn't seem like it was gonna close
In-distribution concept baseline
RD-ID03
Exact training-supported relation × candidate-category baseline
null
431
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 4 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
4
1
Empty
null
null
null
null
null
null
null
between
null
null
null
null
crushed soda can
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
4
4
VLA-0 (Epoch 12)
Failure
Referential Description
33
Pick up the crushed soda can furthest from the green pen and place it in the bowl.
crushed Coke can; crushed Pepsi can; green pen
crushed Pepsi can
turned towards pepsi can sort of, but then started gripping in middle of nowhere between the pepsi can and coke can; drifted closer to coke can until it push the coke can to the left
In-distribution concept baseline
RD-ID04
Exact training-supported relation × candidate-category baseline
null
415
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 3 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
3
1
Empty
null
null
null
null
null
null
null
furthest
null
null
null
null
crushed soda can
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
5
5
VLA-0 (Epoch 12)
Failure
Referential Description
34
Pick up the crushed soda can furthest from the yellow pen and place it in the bowl.
crushed Coke can; crushed Pepsi can; yellow pen; airplane toy with propellers
crushed Coke can
went to coke can in correct gripping strategy, but with grippers closed
In-distribution concept baseline
RD-ID05
Exact training-supported relation × candidate-category baseline
null
416
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 4 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
4
1
Empty
null
null
null
null
null
null
null
furthest
null
null
null
null
crushed soda can
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
6
6
VLA-0 (Epoch 12)
Failure
Referential Description
35
Pick up the pencil sharpener nearest to the crushed Coke can and place it in the bowl.
green pencil sharpener; red pencil sharpener; crushed Coke can
green pencil sharpener
went for red pencil sharpener, but couldn't get it
In-distribution concept baseline
RD-ID06
Exact training-supported relation × candidate-category baseline
null
450
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 3 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
3
1
Empty
null
null
null
null
null
null
null
closest
null
null
null
null
pencil sharpener
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
7
7
VLA-0 (Epoch 12)
Failure
Referential Description
36
Move the pencil sharpener that is closest to the crushed Pepsi can into the bowl.
green pencil sharpener; red pencil sharpener; crushed Pepsi can; airplane toy with propellers
red pencil sharpener
started going to the pepsi can with top down gripping position, then suddenly jerked up to soada can gripping strategy; knocked the soda can out of the workspace
In-distribution concept baseline
RD-ID07
Exact training-supported relation × candidate-category baseline
null
451
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 4 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
4
1
Empty
null
null
null
null
null
null
null
closest
null
null
null
null
pencil sharpener
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
8
8
VLA-0 (Epoch 12)
Failure
Referential Description
37
Move the pencil sharpener that is closest to the green pen into the bowl.
green pencil sharpener; red pencil sharpener; green pen
green pencil sharpener
almost gripped the green sharpener, but after closing gripper, couldn't recover
In-distribution concept baseline
RD-ID08
Exact training-supported relation × candidate-category baseline
null
452
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 3 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
3
1
Empty
null
null
null
null
null
null
null
closest
null
null
null
null
pencil sharpener
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
9
9
VLA-0 (Epoch 12)
Failure
Referential Description
38
Pick up the pencil sharpener furthest from the yellow pen and place it in the bowl.
green pencil sharpener; red pencil sharpener; yellow pen; airplane toy without propellers
green pencil sharpener
went towards teh yellow pen, but couldn't grip it
In-distribution concept baseline
RD-ID09
Exact training-supported relation × candidate-category baseline
null
466
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 4 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
4
1
Empty
null
null
null
null
null
null
null
furthest
null
null
null
null
pencil sharpener
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
10
10
VLA-0 (Epoch 12)
Failure
Referential Description
39
Pick up the pencil sharpener furthest from the green pen and place it in the bowl.
green pencil sharpener; red pencil sharpener; green pen
red pencil sharpener
went for the red pencil sharpener; couldn't pick up
In-distribution concept baseline
RD-ID10
Exact training-supported relation × candidate-category baseline
null
465
Exact manifest cell: prompt, relation × candidate-category combination, target identity, and object membership are explicitly represented in training; physical positions are manually recreated with a clear margin.
1
Place the 3 listed objects. Arrange Expected Target as the unique answer to the prompt; keep the bowl empty.
The relation, candidate category, target identity, and object membership occur together in the cited training episode.
null
3
1
Empty
null
null
null
null
null
null
null
furthest
null
null
null
null
pencil sharpener
null
null
null
Eval_Referential_Disambiguation.xlsm
VLA0 Results
11
11
VLA-0 (Epoch 12)
Failure
Relational Placement
40
Move the green pen to the front of the pink marker.
green pen, pink marker, red block
green pen → 2–8 cm edge gap in front of pink marker; centroid direction within ±45°.
went for red block but didn't even pick it up
In-distribution concept baseline
RP-ID-CF01
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
null
204
null
null
Build the listed neutral scene and save the first frame.
Exact explicitly trained ordered relation-object cell from training episode 204. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
front
green pen
pink marker
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
2
2
VLA-0 (Epoch 12)
Failure
Relational Placement
41
Move the green pen to the back of the pink marker.
green pen, pink marker, red block
green pen → 2–8 cm edge gap behind pink marker; centroid direction within ±45°.
didn't move
In-distribution concept baseline
RP-ID-CF01
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
0
217
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 217. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
behind
green pen
pink marker
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
3
3
VLA-0 (Epoch 12)
Failure
Relational Placement
42
Move the pink marker to the front of the green pen.
green pen, pink marker, red block
pink marker → 2–8 cm edge gap in front of green pen; centroid direction within ±45°.
went for red block again; didn't move it or anything
In-distribution concept baseline
RP-ID-CF01
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
0
305
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 305. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
front
pink marker
green pen
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
4
4
VLA-0 (Epoch 12)
Failure
Relational Placement
43
Move the pink marker to the back of the green pen.
green pen, pink marker, red block
pink marker → 2–8 cm edge gap behind green pen; centroid direction within ±45°.
kind of jittered around home
In-distribution concept baseline
RP-ID-CF01
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
0
317
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 317. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
behind
pink marker
green pen
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
5
5
VLA-0 (Epoch 12)
Failure
Relational Placement
44
Move the green pen to the front of the blue block.
green pen, blue block, pink marker, red block
green pen → 2–8 cm edge gap in front of blue block; centroid direction within ±45°.
went for pink pen but didn't pick it
In-distribution concept baseline
RP-ID-CF02
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
null
208
null
null
Build the listed neutral scene and save the first frame.
Exact explicitly trained ordered relation-object cell from training episode 208. The four adjacent prompts share one initially unsatisfied scene.
null
4
null
null
null
null
null
null
null
null
null
null
front
green pen
blue block
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
6
6
VLA-0 (Epoch 12)
Failure
Relational Placement
45
Move the green pen to the back of the blue block.
green pen, blue block, pink marker, red block
green pen → 2–8 cm edge gap behind blue block; centroid direction within ±45°.
similar trajectory as before
In-distribution concept baseline
RP-ID-CF02
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
4
221
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 221. The four adjacent prompts share one initially unsatisfied scene.
null
4
null
null
null
null
null
null
null
null
null
null
behind
green pen
blue block
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
7
7
VLA-0 (Epoch 12)
Failure
Relational Placement
46
Put the blue block on the table in front of the green pen.
green pen, blue block, pink marker, red block
blue block → 2–8 cm edge gap in front of green pen; centroid direction within ±45°.
went to green pen then returned home
In-distribution concept baseline
RP-ID-CF02
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
4
350
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 350. The four adjacent prompts share one initially unsatisfied scene.
null
4
null
null
null
null
null
null
null
null
null
null
front
blue block
green pen
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
8
8
VLA-0 (Epoch 12)
Failure
Relational Placement
47
Put the blue block on the table behind the green pen.
green pen, blue block, pink marker, red block
blue block → 2–8 cm edge gap behind green pen; centroid direction within ±45°.
didn't move
In-distribution concept baseline
RP-ID-CF02
Four-way trained-cell counterfactual: front/behind polarity × target/reference role
4
362
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 362. The four adjacent prompts share one initially unsatisfied scene.
null
4
null
null
null
null
null
null
null
null
null
null
behind
blue block
green pen
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
9
9
VLA-0 (Epoch 12)
Failure
Relational Placement
48
Place the red block to the left of the pink marker.
red block, pink marker, blue block
red block → 2–8 cm edge gap to the left of pink marker; centroid direction within ±45°.
was gripping around the middle of the workspace; between blue and red block; did move towards the blue block; but couldn't pick it up
In-distribution concept baseline
RP-ID-CF03
Four-way trained-cell counterfactual: left/right polarity × target/reference role
null
275
null
null
Build the listed neutral scene and save the first frame.
Exact explicitly trained ordered relation-object cell from training episode 275. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
left
red block
pink marker
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
10
10
VLA-0 (Epoch 12)
Failure
Relational Placement
49
Place the red block to the right of the pink marker.
red block, pink marker, blue block
red block → 2–8 cm edge gap to the right of pink marker; centroid direction within ±45°.
went for blue block
In-distribution concept baseline
RP-ID-CF03
Four-way trained-cell counterfactual: left/right polarity × target/reference role
8
287
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 287. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
right
red block
pink marker
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
11
11
VLA-0 (Epoch 12)
Failure
Relational Placement
50
Place the pink marker to the left of the red block.
red block, pink marker, blue block
pink marker → 2–8 cm edge gap to the left of red block; centroid direction within ±45°.
went for red block but couldn't pick it
In-distribution concept baseline
RP-ID-CF03
Four-way trained-cell counterfactual: left/right polarity × target/reference role
8
333
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 333. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
left
pink marker
red block
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
12
12
VLA-0 (Epoch 12)
Failure
Relational Placement
51
Place the pink marker to the right of the red block.
red block, pink marker, blue block
pink marker → 2–8 cm edge gap to the right of red block; centroid direction within ±45°.
didn't move
In-distribution concept baseline
RP-ID-CF03
Four-way trained-cell counterfactual: left/right polarity × target/reference role
8
346
null
null
Reset the moved object to its marked pose; preserve the exact scene and change only the instruction.
Exact explicitly trained ordered relation-object cell from training episode 346. The four adjacent prompts share one initially unsatisfied scene.
null
3
null
null
null
null
null
null
null
null
null
null
right
pink marker
red block
null
null
null
null
null
Eval_Relational_Placement (1).xlsm
VLA0
13
13
VLA-0 (Epoch 12)
Failure
Relative Size
52
Move the object that is smallest into the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 red pencil sharpener; P2 green block; P3 juggling ball; P4 crushed Pepsi can. Adjacent centers 8 cm apart; row centered on the workspace; all objects fully visible and grasp paths clear.
MOVE red pencil sharpener to bowl
moved towards the red pencil sharpener, but didn't open gripper to pick it up.
In-distribution concept baseline
SZ-ID-SH01
smallest
null
1005
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1005; this establishes the familiar smallest × red pencil sharpener cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
red pencil sharpener
null
null
null
Left
8
Explicitly trained smallest × red pencil sharpener
Eval_Size_Recognition (1).xlsm
VLA-0 Results
2
2
VLA-0 (Epoch 12)
Failure
Relative Size
53
Pick up the smallest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 crushed Pepsi can; P2 juggling ball; P3 green block. Adjacent centers 12 cm apart; row shifted 4 cm toward robot-left; all objects fully visible and grasp paths clear.
MOVE green block to bowl
also went towards the green block, but didn't go to the top of it; didn't open grippers properly to try and pick it up
In-distribution concept baseline
SZ-ID-GB01
smallest
null
1050
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1050; this establishes the familiar smallest × green block cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
green block
null
null
null
Right
12
Explicitly trained smallest × green block
Eval_Size_Recognition (1).xlsm
VLA-0 Results
3
3
VLA-0 (Epoch 12)
Failure
Relative Size
54
Put the largest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 juggling ball; P2 red pencil sharpener; P3 green block. Adjacent centers 16 cm apart; row shifted 4 cm toward robot-right; all objects fully visible and grasp paths clear.
MOVE juggling ball to bowl
went between the red pencil sharpener and green block, but didn't pick anything up
In-distribution concept baseline
SZ-ID-JB01
largest
null
1125
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1125; this establishes the familiar largest × juggling ball cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
juggling ball
null
null
null
Left
16
Explicitly trained largest × juggling ball
Eval_Size_Recognition (1).xlsm
VLA-0 Results
4
4
VLA-0 (Epoch 12)
Failure
Relative Size
55
Pick up the largest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 red pencil sharpener; P2 green block; P3 juggling ball; P4 crushed Pepsi can. Adjacent centers 8 cm apart; row shifted 4 cm toward the robot; all objects fully visible and grasp paths clear.
MOVE crushed Pepsi can to bowl
went into the gripping strategy for the can, but didn't open the gripper to pick it up
In-distribution concept baseline
SZ-ID-PC01
largest
null
1155
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1155; this establishes the familiar largest × crushed Pepsi can cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
crushed Pepsi can
null
null
null
Right
8
Explicitly trained largest × crushed Pepsi can
Eval_Size_Recognition (1).xlsm
VLA-0 Results
5
5
VLA-0 (Epoch 12)
Failure
Relative Size
56
Pick up the smallest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 juggling ball; P2 crushed Pepsi can; P3 green block; P4 red pencil sharpener. Adjacent centers 12 cm apart; row shifted 4 cm toward robot-left; all objects fully visible and grasp paths clear.
MOVE red pencil sharpener to bowl
went to pepsi can, but with non soda can gripping strategy
In-distribution concept baseline
SZ-ID-SH02
smallest
null
1006
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1006; this establishes the familiar smallest × red pencil sharpener cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
red pencil sharpener
null
null
null
Right
12
Explicitly trained smallest × red pencil sharpener
Eval_Size_Recognition (1).xlsm
VLA-0 Results
6
6
VLA-0 (Epoch 12)
Failure
Relative Size
57
Pick up the smallest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 green block; P2 juggling ball; P3 crushed Pepsi can. Adjacent centers 16 cm apart; row shifted 4 cm toward robot-right; all objects fully visible and grasp paths clear.
MOVE green block to bowl
went to middle of workspace, but didn't grab anything
In-distribution concept baseline
SZ-ID-GB02
smallest
null
1051
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1051; this establishes the familiar smallest × green block cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
green block
null
null
null
Left
16
Explicitly trained smallest × green block
Eval_Size_Recognition (1).xlsm
VLA-0 Results
7
7
VLA-0 (Epoch 12)
Failure
Relative Size
58
Pick up the largest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 green block; P2 red pencil sharpener; P3 juggling ball. Adjacent centers 8 cm apart; row shifted 4 cm toward the robot; all objects fully visible and grasp paths clear.
MOVE juggling ball to bowl
did go to the juggling ball but couldn't grip it
In-distribution concept baseline
SZ-ID-JB02
largest
null
1126
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1126; this establishes the familiar largest × juggling ball cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
juggling ball
null
null
null
Right
8
Explicitly trained largest × juggling ball
Eval_Size_Recognition (1).xlsm
VLA-0 Results
8
8
VLA-0 (Epoch 12)
Failure
Relative Size
59
Move the object that is largest into the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 crushed Pepsi can; P2 green block; P3 juggling ball; P4 red pencil sharpener. Adjacent centers 12 cm apart; row shifted 4 cm away from the robot; all objects fully visible and grasp paths clear.
MOVE crushed Pepsi can to bowl
turned to juggling ball with soda can gripping strategy; didn't grip the juggling ball
In-distribution concept baseline
SZ-ID-PC02
largest
null
1156
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1156; this establishes the familiar largest × crushed Pepsi can cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
crushed Pepsi can
null
null
null
Left
12
Explicitly trained largest × crushed Pepsi can
Eval_Size_Recognition (1).xlsm
VLA-0 Results
9
9
VLA-0 (Epoch 12)
Failure
Relative Size
60
Move the object that is smallest into the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 red pencil sharpener; P2 crushed Pepsi can; P3 green block; P4 juggling ball. Adjacent centers 16 cm apart; row shifted 4 cm toward robot-right; all objects fully visible and grasp paths clear.
MOVE red pencil sharpener to bowl
kind of went for green block
In-distribution concept baseline
SZ-ID-SH03
smallest
null
1007
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1007; this establishes the familiar smallest × red pencil sharpener cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
red pencil sharpener
null
null
null
Left
16
Explicitly trained smallest × red pencil sharpener
Eval_Size_Recognition (1).xlsm
VLA-0 Results
10
10
VLA-0 (Epoch 12)
Failure
Relative Size
61
Put the smallest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 crushed Pepsi can; P2 juggling ball; P3 green block. Adjacent centers 8 cm apart; row shifted 4 cm toward the robot; all objects fully visible and grasp paths clear.
MOVE green block to bowl
went to juggling ball and could not pick up
In-distribution concept baseline
SZ-ID-GB03
smallest
null
1052
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1052; this establishes the familiar smallest × green block cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
green block
null
null
null
Right
8
Explicitly trained smallest × green block
Eval_Size_Recognition (1).xlsm
VLA-0 Results
11
11
VLA-0 (Epoch 12)
Failure
Relative Size
62
Pick up the largest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 juggling ball; P2 red pencil sharpener; P3 green block. Adjacent centers 12 cm apart; row shifted 4 cm away from the robot; all objects fully visible and grasp paths clear.
MOVE juggling ball to bowl
went to red pencil sharpener and did not move
In-distribution concept baseline
SZ-ID-JB03
largest
null
1127
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1127; this establishes the familiar largest × juggling ball cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
juggling ball
null
null
null
Left
12
Explicitly trained largest × juggling ball
Eval_Size_Recognition (1).xlsm
VLA-0 Results
12
12
VLA-0 (Epoch 12)
Failure
Relative Size
63
Put the largest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 juggling ball; P2 red pencil sharpener; P3 green block; P4 crushed Pepsi can. Adjacent centers 16 cm apart; row centered on the workspace; all objects fully visible and grasp paths clear.
MOVE crushed Pepsi can to bowl
went into the can gripping strategy but didn't get it
In-distribution concept baseline
SZ-ID-PC03
largest
null
1157
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1157; this establishes the familiar largest × crushed Pepsi can cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
crushed Pepsi can
null
null
null
Right
16
Explicitly trained largest × crushed Pepsi can
Eval_Size_Recognition (1).xlsm
VLA-0 Results
13
13
VLA-0 (Epoch 12)
Failure
Relative Size
64
Put the smallest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 green block; P2 juggling ball; P3 crushed Pepsi can; P4 red pencil sharpener. Adjacent centers 8 cm apart; row shifted 4 cm toward the robot; all objects fully visible and grasp paths clear.
MOVE red pencil sharpener to bowl
null
In-distribution concept baseline
SZ-ID-SH04
smallest
null
1008
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1008; this establishes the familiar smallest × red pencil sharpener cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
red pencil sharpener
null
null
null
Right
8
Explicitly trained smallest × red pencil sharpener
Eval_Size_Recognition (1).xlsm
VLA-0 Results
14
14
VLA-0 (Epoch 12)
Failure
Relative Size
65
Put the smallest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 green block; P2 juggling ball; P3 crushed Pepsi can. Adjacent centers 12 cm apart; row shifted 4 cm away from the robot; all objects fully visible and grasp paths clear.
MOVE green block to bowl
slowly rotated towards right but didn't do anything
In-distribution concept baseline
SZ-ID-GB04
smallest
null
1053
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1053; this establishes the familiar smallest × green block cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
green block
null
null
null
Left
12
Explicitly trained smallest × green block
Eval_Size_Recognition (1).xlsm
VLA-0 Results
15
15
VLA-0 (Epoch 12)
Failure
Relative Size
66
Pick up the largest item and place it in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 green block; P2 red pencil sharpener; P3 juggling ball. Adjacent centers 16 cm apart; row centered on the workspace; all objects fully visible and grasp paths clear.
MOVE juggling ball to bowl
went for the green block, but couldn't grab it
In-distribution concept baseline
SZ-ID-JB04
largest
null
1128
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1128; this establishes the familiar largest × juggling ball cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
juggling ball
null
null
null
Right
16
Explicitly trained largest × juggling ball
Eval_Size_Recognition (1).xlsm
VLA-0 Results
16
16
VLA-0 (Epoch 12)
Failure
Relative Size
67
Put the largest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 crushed Pepsi can; P2 red pencil sharpener; P3 green block; P4 juggling ball. Adjacent centers 8 cm apart; row shifted 4 cm toward robot-left; all objects fully visible and grasp paths clear.
MOVE crushed Pepsi can to bowl
didn't rotate far enough to the right to grab the soda can. Had the right gripping strategy but was going to pencil sharpener
In-distribution concept baseline
SZ-ID-PC04
largest
null
1158
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1158; this establishes the familiar largest × crushed Pepsi can cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
crushed Pepsi can
null
null
null
Left
8
Explicitly trained largest × crushed Pepsi can
Eval_Size_Recognition (1).xlsm
VLA-0 Results
17
17
VLA-0 (Epoch 12)
Failure
Relative Size
68
Move the object that is smallest into the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 red pencil sharpener; P2 juggling ball; P3 crushed Pepsi can; P4 green block. Adjacent centers 12 cm apart; row shifted 4 cm away from the robot; all objects fully visible and grasp paths clear.
MOVE red pencil sharpener to bowl
went for juggling ball, then aggressively lifted wrist up and rolled it out of the workspace
In-distribution concept baseline
SZ-ID-SH05
smallest
null
1009
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1009; this establishes the familiar smallest × red pencil sharpener cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
red pencil sharpener
null
null
null
Left
12
Explicitly trained smallest × red pencil sharpener
Eval_Size_Recognition (1).xlsm
VLA-0 Results
18
18
VLA-0 (Epoch 12)
Failure
Relative Size
69
Put the smallest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 crushed Pepsi can; P2 juggling ball; P3 green block. Adjacent centers 16 cm apart; row centered on the workspace; all objects fully visible and grasp paths clear.
MOVE green block to bowl
moved to the green block, but didn't open up the gripper to pick it up
In-distribution concept baseline
SZ-ID-GB05
smallest
null
1054
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1054; this establishes the familiar smallest × green block cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
green block
null
null
null
Right
16
Explicitly trained smallest × green block
Eval_Size_Recognition (1).xlsm
VLA-0 Results
19
19
VLA-0 (Epoch 12)
Failure
Relative Size
70
Put the largest object in the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 juggling ball; P2 red pencil sharpener; P3 green block. Adjacent centers 8 cm apart; row shifted 4 cm toward robot-left; all objects fully visible and grasp paths clear.
MOVE juggling ball to bowl
tried to pick the red pencil sharpener and couldn't
In-distribution concept baseline
SZ-ID-JB05
largest
null
1129
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1129; this establishes the familiar largest × juggling ball cell.
null
3
null
null
null
null
null
null
null
null
null
null
null
juggling ball
null
null
null
Left
8
Explicitly trained largest × juggling ball
Eval_Size_Recognition (1).xlsm
VLA-0 Results
20
20
VLA-0 (Epoch 12)
Failure
Relative Size
71
Move the object that is largest into the bowl.
Bowl: empty. Table, robot-left→robot-right: P1 green block; P2 juggling ball; P3 red pencil sharpener; P4 crushed Pepsi can. Adjacent centers 12 cm apart; row shifted 4 cm toward robot-right; all objects fully visible and grasp paths clear.
MOVE crushed Pepsi can to bowl
moved between the pepsi can and juggling ball with soda can gripping technique.
In-distribution concept baseline
SZ-ID-PC05
largest
null
1159
null
null
Build the listed manifest-supported scene.
Exact prompt and object composition from training episode 1159; this establishes the familiar largest × crushed Pepsi can cell.
null
4
null
null
null
null
null
null
null
null
null
null
null
crushed Pepsi can
null
null
null
Right
12
Explicitly trained largest × crushed Pepsi can
Eval_Size_Recognition (1).xlsm
VLA-0 Results
21
21
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
0
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [blue pen, yellow block, green pencil sharpener]
ADD any one table object; finish with 1 object in the bowl
picked the yellow block
In-distribution concept baseline
ID-F01
In-distribution: one via trained 0→1 addition
null
801
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 801.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
2
2
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
1
Adjust the bowl so that it contains two objects.
Bowl objects: [yellow block] | Table objects: [blue pen, green pencil sharpener]
ADD any one table object; finish with 2 objects in the bowl
didn't move
In-distribution concept baseline
ID-F01
In-distribution: two via trained 1→2 addition
null
809
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 809.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
3
3
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
2
Leave exactly three objects in the bowl.
Bowl objects: [blue pen, yellow block, green pencil sharpener, airplane toy without propellers] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
picked both the airplane toy and green pencil sharpener simulataneously so left two items int he bowl
In-distribution concept baseline
ID-F01
In-distribution: three via trained 4→3 removal
null
843
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 843.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
4
4
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
3
Adjust the bowl so that it contains zero objects.
Bowl objects: [blue pen] | Table objects: [green pencil sharpener, airplane toy without propellers]
REMOVE the only bowl object; finish with 0 objects in the bowl
null
In-distribution concept baseline
ID-F01
In-distribution: zero via trained 1→0 removal
null
826
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 826.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F01
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
5
5
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
4
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [yellow block, blue pen]
ADD any one table object; finish with 1 object in the bowl
null
In-distribution concept baseline
ID-F02
In-distribution: one via trained 0→1 addition
null
852
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 852.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
6
6
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
5
Adjust the bowl so that it contains two objects.
Bowl objects: [green pencil sharpener] | Table objects: [yellow block, airplane toy without propellers]
ADD any one table object; finish with 2 objects in the bowl
tried to pick yellow block but couldn't
In-distribution concept baseline
ID-F02
In-distribution: two via trained 1→2 addition
null
859
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 859.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
7
7
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
6
Leave exactly three objects in the bowl.
Bowl objects: [yellow block, green pencil sharpener, airplane toy without propellers, blue pen] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
went to bowl, but couldn't grab anything
In-distribution concept baseline
ID-F02
In-distribution: three via trained 4→3 removal
null
893
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 893.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
8
8
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
7
Adjust the bowl so that it contains zero objects.
Bowl objects: [yellow block] | Table objects: [blue pen]
REMOVE the only bowl object; finish with 0 objects in the bowl
null
In-distribution concept baseline
ID-F02
In-distribution: zero via trained 1→0 removal
null
878
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 878.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F02
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
9
9
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
8
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [green pencil sharpener, yellow block]
ADD any one table object; finish with 1 object in the bowl
null
In-distribution concept baseline
ID-F03
In-distribution: one via trained 0→1 addition
null
902
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 902.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F03
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
10
10
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
9
Adjust the bowl so that it contains two objects.
Bowl objects: [blue pen] | Table objects: [green pencil sharpener, airplane toy without propellers]
ADD any one table object; finish with 2 objects in the bowl
went for the airplane toy 5 times, but never moved it.
In-distribution concept baseline
ID-F03
In-distribution: two via trained 1→2 addition
null
909
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 909.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F03
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
11
11
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
10
Leave exactly three objects in the bowl.
Bowl objects: [green pencil sharpener, airplane toy without propellers, blue pen, yellow block] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
went to bowl but couldn't pick anything up
In-distribution concept baseline
ID-F03
In-distribution: three via trained 4→3 removal
null
943
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 943.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F03
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
12
12
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
11
Make the bowl contain exactly zero objects.
Bowl objects: [green pencil sharpener] | Table objects: [yellow block]
REMOVE the only bowl object; finish with 0 objects in the bowl
null
In-distribution concept baseline
ID-F03
In-distribution: zero via trained 1→0 removal
null
928
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 928.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F03
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
13
13
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
12
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [airplane toy without propellers, blue pen, yellow block]
ADD any one table object; finish with 1 object in the bowl
null
In-distribution concept baseline
ID-F04
In-distribution: one via trained 0→1 addition
null
951
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 951.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F04
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
14
14
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
13
Make the bowl contain exactly two objects.
Bowl objects: [green pencil sharpener] | Table objects: [airplane toy without propellers, yellow block]
ADD any one table object; finish with 2 objects in the bowl
null
In-distribution concept baseline
ID-F04
In-distribution: two via trained 1→2 addition
null
960
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 960.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F04
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
15
15
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
14
Leave exactly three objects in the bowl.
Bowl objects: [airplane toy without propellers, blue pen, yellow block, green pencil sharpener] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
went to bowl but couldn't remove anything
In-distribution concept baseline
ID-F04
In-distribution: three via trained 4→3 removal
null
993
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 993.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F04
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
16
16
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
15
Leave no objects in the bowl.
Bowl objects: [airplane toy without propellers] | Table objects: [green pencil sharpener]
REMOVE the only bowl object; finish with 0 objects in the bowl
null
In-distribution concept baseline
ID-F04
In-distribution: zero via trained 1→0 removal
null
978
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 978.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F04
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
17
17
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
16
Leave exactly one object in the bowl.
Bowl objects: [none] | Table objects: [yellow block]
ADD the only table object; finish with 1 object in the bowl
null
In-distribution concept baseline
ID-F05
In-distribution: one via trained 0→1 addition
null
857
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 857.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F05
0
1
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
18
18
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
17
Make the bowl contain exactly two objects.
Bowl objects: [green pencil sharpener] | Table objects: [yellow block]
ADD the only table object; finish with 2 objects in the bowl
didn't move
In-distribution concept baseline
ID-F05
In-distribution: two via trained 1→2 addition
null
865
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 865.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F05
1
2
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
19
19
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
18
Leave exactly three objects in the bowl.
Bowl objects: [airplane toy without propellers, blue pen, yellow block, green pencil sharpener] | Table objects: [none]
REMOVE any one bowl object; finish with 3 objects in the bowl
null
In-distribution concept baseline
ID-F05
In-distribution: three via trained 4→3 removal
null
994
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 994.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F05
4
3
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
20
20
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
19
Make the bowl contain exactly zero objects.
Bowl objects: [yellow block] | Table objects: [none]
REMOVE the only bowl object; finish with 0 objects in the bowl
null
In-distribution concept baseline
ID-F05
In-distribution: zero via trained 1→0 removal
null
880
Exact trained cell: prompt, count transition, and scene-object membership match; physical layout is not copied exactly.
1
Recreate the listed bowl/table membership from training episode 880.
Exact trained count transition and scene-object membership. Any object that reaches the requested final count is acceptable.
null
null
null
null
null
null
ID-F05
1
0
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
21
21
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
20
Leave exactly one object in the bowl.
Bowl objects: [blue pen, airplane toy without propellers] | Table objects: [yellow block, green pencil sharpener]
REMOVE any one bowl object; finish with 1 object in the bowl
didn't move
Concept/compositional transfer
CT-CF09
Primary: one via held-out 2→1 removal
null
831
Representative direct-instruction trajectory for blue pen; the evaluated count × transition composition is intentionally absent.
3
Build the listed count-2 scene.
The goal word one is trained, and this 2→1 trajectory is trained under a direct instruction, but the combination is withheld.
null
null
null
null
null
null
CT-CF09-REMOVE
2
1
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
22
22
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
21
Move the airplane toy without propellers from the bowl to the table.
Bowl objects: [blue pen, airplane toy without propellers] | Table objects: [yellow block, green pencil sharpener]
REMOVE airplane toy without propellers
null
Matched manipulation control
CT-CF09
Matched direct REMOVE control
20
985
Exact direct-prompt trajectory for the named object and action direction; evaluation layout is reset manually.
3
Reset the exact count-2 scene; change only the prompt.
The direct prompt is present in Counting training for this same-task object and action direction.
null
null
null
null
null
null
CT-CF09-REMOVE
2
1
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
23
23
GR00T-N1.7 (Frozen LLM, Epoch 12)
Success
Counting
22
Move the green pencil sharpener into the bowl.
Bowl objects: [blue pen, airplane toy without propellers] | Table objects: [yellow block, green pencil sharpener]
ADD green pencil sharpener
null
Matched manipulation control
CT-CF09
Matched direct ADD control
null
917
Exact direct-prompt trajectory for the named object and action direction; evaluation layout is reset manually.
3
Reset the exact count-2 scene; change only the prompt.
The direct prompt is present in Counting training for this same-task object and action direction.
null
null
null
null
null
null
CT-CF09-ADD
2
3
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
24
24
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
23
Leave exactly three objects in the bowl.
Bowl objects: [blue pen, airplane toy without propellers] | Table objects: [yellow block, green pencil sharpener]
ADD any one table object; finish with 3 objects in the bowl
didn't move
Concept/compositional transfer
CT-CF09
Primary: three via held-out 2→3 addition
22
916
Representative direct-instruction trajectory for green pencil sharpener; the evaluated count × transition composition is intentionally absent.
3
Reset the exact count-2 scene; change only the prompt.
The goal word three is trained, and this 2→3 trajectory is trained under a direct instruction, but the combination is withheld.
null
null
null
null
null
null
CT-CF09-ADD
2
3
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
25
25
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
24
Leave exactly three objects in the bowl.
Bowl objects: [green pencil sharpener, airplane toy without propellers] | Table objects: [blue pen, yellow block]
ADD any one table object; finish with 3 objects in the bowl
didn't move
Concept/compositional transfer
CT-CF24
Primary: three via held-out 2→3 addition
null
816
Representative direct-instruction trajectory for blue pen; the evaluated count × transition composition is intentionally absent.
4
Reset the exact count-2 scene; change only the prompt.
The goal word three is trained, and this 2→3 trajectory is trained under a direct instruction, but the combination is withheld.
null
null
null
null
null
null
CT-CF24-ADD
2
3
ADD
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
26
26
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
25
Leave exactly one object in the bowl.
Bowl objects: [green pencil sharpener, airplane toy without propellers] | Table objects: [blue pen, yellow block]
REMOVE any one bowl object; finish with 1 object in the bowl
didn't move
Concept/compositional transfer
CT-CF24
Primary: one via held-out 2→1 removal
24
981
Representative direct-instruction trajectory for airplane toy without propellers; the evaluated count × transition composition is intentionally absent.
4
Build the listed count-2 scene.
The goal word one is trained, and this 2→1 trajectory is trained under a direct instruction, but the combination is withheld.
null
null
null
null
null
null
CT-CF24-REMOVE
2
1
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
27
27
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
26
Leave exactly one object in the bowl.
Bowl objects: [blue pen, yellow block] | Table objects: [green pencil sharpener]
REMOVE any one bowl object; finish with 1 object in the bowl
didn't move
Concept/compositional transfer
CT-CF01
Primary: one via held-out 2→1 removal
null
831
Representative direct-instruction trajectory for blue pen; the evaluated count × transition composition is intentionally absent.
2
Build the listed count-2 scene.
The goal word one is trained, and this 2→1 trajectory is trained under a direct instruction, but the combination is withheld.
null
null
null
null
null
null
CT-CF01-REMOVE
2
1
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
28
28
GR00T-N1.7 (Frozen LLM, Epoch 12)
Failure
Counting
27
Move the blue pen from the bowl to the table.
Bowl objects: [blue pen, yellow block] | Table objects: [green pencil sharpener]
REMOVE blue pen
went to bowl but couldn't take the pen out
Matched manipulation control
CT-CF01
Matched direct REMOVE control
26
835
Exact direct-prompt trajectory for the named object and action direction; evaluation layout is reset manually.
2
Reset the exact count-2 scene; change only the prompt.
The direct prompt is present in Counting training for this same-task object and action direction.
null
null
null
null
null
null
CT-CF01-REMOVE
2
1
REMOVE
null
null
null
null
null
null
null
null
null
null
Eval_Counting.xlsm
GR00T vision only Results
29
29
End of preview. Expand in Data Studio

Spa-Bench data information

This repository exposes two reviewer-friendly tables through Hugging Face Data Studio:

  • Model Evaluation Data Information: 2,892 scored physical rollouts across five checkpoints.
  • Training Data Information: the 1,200-row prompt and compositional-design manifest.

Use the configuration selector in the viewer to switch between the two tables. Task names are normalized to Counting, Ordinal Position, Physical State, Referential Description, Relational Placement, and Relative Size.

Training-design summary

All atomic concepts and object identities appear during training. The benchmark withholds selected concept–argument combinations rather than unseen words, objects, or motor outputs. Each task family contains 200 demonstrations.

Models

Evaluation rollouts

Training dataset

Double-blind release note

Author, institution, source-repository, and archival citation details are intentionally omitted during review. They will be restored in the archival release.

Downloads last month
42