MimicGen & DexMimicGen Reward Modeling
Collection
Causal SARM / ReWiND reward models vs failure-aware FARM: MimicGen Coffee, Mug Cleanup; DexMimicGen Threading, Three-Piece (data + checkpoints) • 23 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
episode: int64
seed: int64
success: bool
steps: int64
n_frames: int64
rollout_exception: bool
npz: string
video: string
n_failure: int64
rollout_dir: string
type: string
max_window: int64
ckpt: string
window_rule: string
episodes: struct<success: list<item: int64>, failure: list<item: int64>>
child 0, success: list<item: int64>
child 0, item: int64
child 1, failure: list<item: int64>
child 0, item: int64
name: string
end_scores: string
window: struct<sampling: string, n_recent: int64>
child 0, sampling: string
child 1, n_recent: int64
n_success: int64
task: string
bootstrap_repo: string
scorer: string
created: timestamp[s]
rollout_manifest: string
to
{'name': Value('string'), 'ckpt': Value('string'), 'type': Value('string'), 'task': Value('string'), 'bootstrap_repo': Value('string'), 'window': {'sampling': Value('string'), 'n_recent': Value('int64')}, 'max_window': Value('int64'), 'window_rule': Value('string'), 'rollout_dir': Value('string'), 'rollout_manifest': Value('string'), 'n_success': Value('int64'), 'n_failure': Value('int64'), 'episodes': {'success': List(Value('int64')), 'failure': List(Value('int64'))}, 'end_scores': Value('string'), 'scorer': Value('string'), 'created': Value('timestamp[s]')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
episode: int64
seed: int64
success: bool
steps: int64
n_frames: int64
rollout_exception: bool
npz: string
video: string
n_failure: int64
rollout_dir: string
type: string
max_window: int64
ckpt: string
window_rule: string
episodes: struct<success: list<item: int64>, failure: list<item: int64>>
child 0, success: list<item: int64>
child 0, item: int64
child 1, failure: list<item: int64>
child 0, item: int64
name: string
end_scores: string
window: struct<sampling: string, n_recent: int64>
child 0, sampling: string
child 1, n_recent: int64
n_success: int64
task: string
bootstrap_repo: string
scorer: string
created: timestamp[s]
rollout_manifest: string
to
{'name': Value('string'), 'ckpt': Value('string'), 'type': Value('string'), 'task': Value('string'), 'bootstrap_repo': Value('string'), 'window': {'sampling': Value('string'), 'n_recent': Value('int64')}, 'max_window': Value('int64'), 'window_rule': Value('string'), 'rollout_dir': Value('string'), 'rollout_manifest': Value('string'), 'n_success': Value('int64'), 'n_failure': Value('int64'), 'episodes': {'success': List(Value('int64')), 'failure': List(Value('int64'))}, 'end_scores': Value('string'), 'scorer': Value('string'), 'created': Value('timestamp[s]')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
DexMimicGen held-out rollouts for reward-model evaluation
100 fresh rollouts per task of the frozen base diffusion policies
(chomeed/dexmg_two_arm_threading,
chomeed/dexmg_two_arm_three_piece_assembly) on new
environment seeds 20000-20099. No reward model was trained on them. They are the evaluation set behind the DexMimicGen
detection results comparing the ReWiND-state baseline with FARM (failure-aware reward model) under a causal window. Caveat:
the Threading rollouts were also used to choose among FARM variants.
| task | episodes | successes | failures | max steps |
|---|---|---|---|---|
| two_arm_threading | 100 | 57 | 43 | 500 |
| two_arm_three_piece_assembly | 100 | 63 | 37 | 500 |
Layout
<task>/manifest.json generation settings (policy checkpoint, seed rule, stop rule, host) + per-episode list
<task>/episodes/epXXXX_seedYYYY.npz agentview, robot0_eye_in_hand, robot1_eye_in_hand: (T+1, 84, 84, 3) uint8;
state (T+1, 18); action; success; seed; steps; rollout_exception; init_sim_state
<task>/episodes/epXXXX_seedYYYY.json episode summary (seed, success, steps, n_frames)
<task>/videos/epXXXX_seedYYYY_<outcome>.mp4 3-camera strip of the stored frames
<task>/frame_check*.{json,png} check that rendered frames match the LeRobot training data (camera, orientation, colour)
scores/<task>/<model>/trajectories.npz per-frame causal rewards, keys success|failure/<episode>/predicted_reward
scores/<task>/<model>/summary.json
Frame 0 is the post-reset observation; frame k is the observation after action k (n_frames = steps + 1). Episode k uses seed 20000 + k, which seeds the object layout, the robot's initial noise and the policy's diffusion noise. Each rollout ran until success or 500 steps. Rollouts continued until there were at least 100 episodes with at least 30 successes and 30 failures.
scores/ holds the two reward models compared in the report. rewind_base_050000 is the ReWiND-state baseline
(chomeed/rewind_state_dexmg_two_arm_<task>).
farm_v7_005000 is FARM v7 at step 5,000
(chomeed/farm_rewind_state_dexmg_two_arm_<task>).
Both are scored per frame with the causal hybrid8 window: 8 frames uniform over the past plus the 8 most recent frames, all
at or before t. Frames 0-14 come from one model pass over frames 0-15, so they are not strictly causal; the report drops
them. Episode score = mean of the last 3 predictions; success is predicted when the score is at least 0.5.
Part of the MimicGen & DexMimicGen Reward Modeling collection.
- Downloads last month
- 52