RoboCost-raw — Raw generator outputs (sim_state + verification evidence)

The upstream raw material for the RoboCost A/B/C datasets — not a packed LeRobot dataset. 93 task-level directories, each holding 30 scenes × safe/unsafe variants for one task at one difficulty level.

This repository is not optional. Two things are impossible without it:

1. Re-deriving costs. reflag.py / cost_v2.py read the per-frame MuJoCo sim_state stored inside <tag>.traj.npz (4975 files, 300 MB total). Those files are not present in the three LeRobot datasets. With them, costs can be recomputed on CPU alone (0.13 s per episode, about 8 minutes for the full set) without re-rendering any video. This is exactly what item 1 of docs/TODO.md in the project repository depends on.
2. Verification evidence. Claims in CHECK_LIST.md such as "27 of 30 levels have a safe-side success rate ≥90%" are backed by the 60 mp4 files and the _summary.log in each directory. The packed datasets do not contain them.

Contents

Type	Count	Description
*.traj.npz	4,975	per-frame MuJoCo sim_state — the only source for re-deriving costs
*.mp4	5,275	side-by-side dual camera (left = agentview, right = wrist), for human review
*.steps.jsonl	5,260	per-step criteria (contact pairs, distances, cost predicates)
*.meta.json	5,262	per-episode metadata (task / level / scene / variant / outcome)
_summary.log	30	success-rate summary table at the top, followed by per-episode JSON criteria
Total	20,825 files	2.1 GB

Directory naming

<class><index>_<object>_L<level>, e.g. C1_bowl_L0, A1_knife_L1, B4_bowl_L0. Inside each directory: scene000..scene029 × {safe, unsafe}. Some directories carry a _v2/_v3 suffix — these are older iterations, and the suffix-free one is authoritative (CHECK_LIST.md in the project repository records which is final).

Usage

hf download greenery8848/robocost-raw --repo-type dataset --local-dir $DATA_ROOT/data/rollouts/libero/gen

# Re-derive the A-class costs (pure CPU)
python scripts/hpc/reflag.py --gen $GEN_ROOT --out $DATA_ROOT/anno/flags_v3.jsonl

reflag.py looks for $GEN/<dir>/<tag>.traj.npz by default, so keep the directory structure as-is. (Note: reflag.py currently crashes on classes B and C — see item 1 of docs/TODO.md.)

Relationship to the three datasets

	Contents	Purpose
robocost-A / -B / -C	LeRobot v2.1 superset (parquet + mp4 + meta)	training, evaluation
robocost-raw (this repo)	raw generator output including sim_state	re-deriving costs, human review, traceability

License and redistribution

Derived from LIBERO-Safety; the upstream repository has no LICENSE file. Contact its authors before redistributing externally.