File size: 3,677 Bytes
54cab51 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 | ---
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- robotics
- vision-language-action
- semantic-reasoning
- compositional-generalization
pretty_name: Reasoning Benchmark 800
---
# Reasoning Benchmark 800
`Reasoning_Benchmark_800` is an 800-episode LeRobot v3 training dataset for
studying semantic and compositional generalization in vision-language-action
models. It combines four completed 200-episode datasets while preserving the
original observations, actions, timing, and videos. Only the training-language
metadata was reassigned.
The experimental control is that held semantic concepts are absent from the
training prompts even though the corresponding visual and motor trajectories
are present. Those trajectories receive an object-identity prompt or a more
generic relation/action prompt. This makes held-prompt evaluation test semantic
generalization rather than unseen motor behavior.
## Composition
| Combined episodes | Task family | Source dataset | Seen-language episodes | Masked/held-support episodes |
|---:|---|---|---:|---:|
| 0–199 | Relational placement | `SP_Relational_Placement_200` | 100 front/back | 100 left/right trajectories prompted as next-to |
| 200–399 | Ordering/sequencing | `SP_Sequencing_200` | 76 seen combinations | 124 first-from-right/third-from-left trajectories prompted by object identity |
| 400–599 | Counting | `SP_Counting_200` | 80 counts 0/2 | 120 counts 1/3 prompted as generic object transfers |
| 600–799 | Size recognition | `MT_Size_Recognition_200` | 150 seen combinations | 50 held object–size combinations prompted by object identity |
This version intentionally excludes referential disambiguation and state
recognition. They can be added in a later benchmark release after those source
datasets are finalized.
## Dataset summary
- 800 episodes
- 413,545 frames
- 30 FPS
- SO follower robot
- 5 RGB video observations: wrist, middle, above, right, and left
- 6-dimensional robot state and action
- 434 unique paraphrased training prompts
- 191 MP4 files, copied without re-encoding
## Prompt assignment
The LeRobot task for each episode is taken only from the final `prompt` column
of `VLA Prompts.xlsx`. The training prompts are represented conventionally by
`meta/tasks.parquet` and the per-frame `task_index` column.
The `benchmark_metadata/` directory contains:
- `prompt_manifest.csv`: compact episode-to-prompt mapping
- `prompt_manifest.jsonl`: full audit mapping, including workbook annotations
- `prompt_manifest_summary.json`: counts and manifest/workbook checksums
- `source_datasets.json`: exact source repositories, revisions, and offsets
- `validation_report.json`: completed integrity checks
The audit JSONL is provenance metadata, not a model input. Training pipelines
should consume the standard LeRobot task field only. See
`benchmark_metadata/TRAINING_NOTES.md`.
## Validation
The merged artifact passed checks for:
- contiguous episode indices 0–799
- contiguous global frame indices 0–413,544
- contiguous per-episode frame indices
- exact episode prompt agreement with the 800-row manifest
- exact per-frame task-index agreement with each episode prompt
- preservation of every source episode's frame count
- SHA-256 multiset equality between all source and merged video files
- zero detected held-concept leakage in the assigned training prompts
## Intended evaluation
Train each architecture on the same 800 episodes and prompt assignments, then
evaluate with separate:
1. matched identity prompts,
2. seen semantic prompts, and
3. held semantic prompts.
Keep held evaluation instructions outside the training input pipeline.
|