ghkim-rlwrld/robodojo-hlp-local-sttp

RoboDojo HLP Local-STTP — v3 training set

One question per row: has the robot moved past the subtask it was given, and if so, what comes next? A high-level planner is called every 16 simulator steps, sees a few head-camera frames and the sentence the controller is currently holding, and answers with a flag plus the next subtask sentence.

41,530 rows built from 995 expert demonstration episodes across 10 RoboDojo tasks. Derived from RoboDojo-Benchmark/RoboDojo; see the licence note at the end.

rows.jsonl   41,530 rows   7,984 transition (flag 1) · 33,546 hold (flag 0)
frames/      <task>/ep<N>/<step>.png — every frame the rows cite, 21 GB
labels/      the subtask annotation the rows were built from, plus camera calibration

A row

{
  "messages": [
    {"role": "system",    "content": "<fixed; defines the output contract>"},
    {"role": "user",      "content": "The task goal is: Stack the three bowls together.\nThe current subtask accepted by the controller is: pick up the bowl at <|box_start|>(775,362)<|box_end|>\nHere are the selected frames ... of particular importance:[<image>]\nHere are the recent frames sampled from the current subtask through CURRENT at stride 24: [<image>, <image>, <image>, <image>, <image>, <image>]\n\nFirst indicate transition with 1 or 0, then predict the subtask and keyframe positions."},
    {"role": "assistant", "content": "1\n{\"current_subtask\": \"pick up the bowl at <bbox>\", \"keyframe_positions\": [6]}"}
  ],
  "images": ["/…/hlp_frames/stack_bowls/ep3000/000000.png", "…"],   // in <image> order
  "objects": {"ref": [], "bbox": [[158, 192]]},                     // [[y, x]], 0–1000
  "metadata": {"task": "stack_bowls", "source_episode_index": 3000, "call_index": 7,
               "current_step": 112, "transition": true,
               "accepted_segment_before_call": 0, "oracle_segment_at_current": 1,
               "active_subtask_start": 0, "recent_steps": [0,24,48,72,96,112],
               "candidate_keyframes": [106], "memory_count": 1}
}

Where the context comes from

part of the prompt	what it is	where to read it
goal sentence	the task's one global instruction, fixed per task	first line of the user turn
current subtask	the sentence the controller is holding, with its point	second line, <|box_start|>(x,y)<|box_end|>
keyframes (past memory)	frames committed at earlier transitions	first metadata.memory_count entries of images
current observation	the segment so far at stride 24, current frame always last	the rest of images; steps in metadata.recent_steps

The answer

first token        1 = the robot has moved past the accepted subtask · 0 = hold
current_subtask    flag 1 → the action to execute now
                   flag 0 → the action that comes AFTER the accepted one (a look-ahead)
                   "unpredictable" when that later action cannot be determined yet
keyframe_positions 1-indexed positions in the RECENT image list nearest to memory keyframes
                   not already held

Images per row: min 1, median 8, max 28. memory_count reaches 14.

Three things that will silently break a consumer

Coordinate order is reversed between the two channels. The accepted-subtask line in the user turn writes (x, y); the box the assistant emits and objects.bbox are (y, x). Same point, numbers swapped. Feeding a model's own output straight back into the accepted line puts y where every row has x — with no error.

The assistant's text carries no digits. The coordinate is the placeholder <bbox>; the value lives in objects.bbox. Scoring the string alone cannot measure coordinate accuracy.

keyframe_positions indexes the recent list, not images. The memory frames come first in images and are not counted in those positions.

Tasks

Full per-task breakdown — goal sentence, subtask vocabulary, counts — in TASKS.md.

task	episodes	transition	hold	img p50
imitate_sorting_sequence	95	941	6,647	16
play_tic_tac_toe	100	1,484	4,637	9
classify_objects	100	1,067	3,759	8
sort_nesting_dolls_by_size	100	841	3,228	8
arrange_largest_number	100	732	3,249	8
cover_blocks	100	1,000	2,828	8
hang_mugs	100	520	3,045	8
stack_bowls	100	461	2,582	7
make_kong	100	419	2,170	7
stack_blocks	100	519	1,401	6

995 episodes, not 1,000: imitate_sorting_sequence is missing joint indices 1217, 1230, 1260, 1268 and 1274, which have no usable annotation.

Provenance

This is the set that trained the adapter the ten-task evaluation actually ran:

train_v3_A.sbatch:109   cat rows_point_v3.jsonl rows_reasoning_v3.jsonl > rows_all_v3.jsonl
train_v3_A.sbatch:89    MODEL_OUTPUT_DIR=.../robodojo_hlp_v3a_lora

Earlier cycles (v1, v2) and unused variants exist in the build directory and are not here. v3 differs from v2 in one label change: the terminal sentence return to the origin position is no longer named as a hold target. That change did not reduce how often the trained planner emits it at run time — measured 19.0 % → 20.4 % on classify_objects — so treat the terminal sentence as an open problem rather than a fixed one.

No validation split

All 995 episodes are in training. The benchmark metric is simulator rollout success, so held-out rows were never built. Anyone measuring generalisation has to carve a split and rebuild.

Licence

Upstream RoboDojo states its licence in two places that disagree: the repository LICENSE file is MIT, while its README.md names a "RoboDojo Non-Commercial Research License" limiting use to non-commercial research, education and evaluation. This derivative is shared for non-commercial research only, which is inside either reading. Commercial use has to be settled with the upstream maintainers first.