RoboCost-C — Obstacle Avoidance Dataset (static obstacles and human-held obstacles)

Simulated manipulation rollouts with per-frame safety cost ground truth — the largest of the three subsets. Safety semantics = avoid the obstacles: it is unsafe when the robot body or the object it carries touches a tabletop obstacle (C1, static/fragile objects) or an object held by a human hand (C2, a drifting hand).

Scale

Item	Value
Episodes	1800 (5 tasks × L0/L1/L2 × 30 scenes × safe/unsafe variants)
Frames	704,091
Frame rate	20 fps
Task success rate on the safe side	855/900
Robot	Franka Panda (LIBERO/robosuite)
Views	agentview + wrist, 256×256 each
Format	LeRobot v2.1 superset

Tasks

1. put the white yellow mug in the microwave and close it
2. pick up the book and place it in the back compartment of the caddy
3. put both moka pots on the stove
4. pick up the akita black bowl on the stove and place it on the plate
5. put both the alphabet soup and the cream cheese box in the basket

Obstacle difficulty ramps up across the three levels L0/L1/L2. The C2 variants replace the static obstacle with a mocap-driven, drifting human hand holding an object.

Schema

Identical to RoboCost-A / RoboCost-B: cost int8[9] + cost_any int8[1], defined in meta/cost_dims.json, with the intended variant labels in meta/variant_modes.json. All three subsets can be merged directly.

Verification status

Of the 30 task-level combinations, 27 have a safe-side success rate ≥90%; 3 sit at 83% (C2 drifting hand at L1: moka 25/30, soup 23/30, mug 25/30 — the tail of the random drift, with safe/unsafe semantics never inverted). The per-level verification record is in CHECK_LIST.md in the project repository; the 60 mp4 files and _summary.log for each task-level directory live in robocost-raw.

⚠️ Known defects

0. meta/episodes_stats.jsonl and stats.json are missing, so the official LeRobot loader cannot open this dataset. The packer never computed statistics. "LeRobot v2.1 superset" refers to the directory layout and the fields, not to out-of-the-box loadability. For now, read the parquet directly with pyarrow (see "Loading" below).

1. cost_v2 has not been re-derived yet. The A class has (see anno/flags_v2.jsonl in the project repository); B and C are waiting on c_cost.build_ctx_c being wired in. The seven instrumentation defects listed on the A card apply here too, and the two that matter most for this subset are:

   • Missing displacement attribution: obstacles that move on their own (the drifting hand, the moving toy) show large displacement even with zero robot contact — the median for toy_train_1 reaches 71 cm. cost_v2 fixes this by re-defining attribution as "stationary for ≥5 frames, then starts moving", which naturally excludes script-driven obstacles because they are already moving at frame 0.
   • Schema drift: cost_dims_fired is a {dimension: step count} dict in A and B but a plain list in C. The field also contains the informational dimension grasp_allowed, which makes bool(cost_dims_fired) the wrong criterion — use cost_any.

2. 20 Hz sampling misses grazing collisions: there really are cases where a milk carton was knocked over (displaced 13.1 cm, fell 7.8 cm, plainly visible in the video) yet zero contact frames were recorded.

3. Per-frame sim_state was written alongside the data, so costs can be recomputed by replaying on CPU alone (about 0.13 s per episode) — no video re-rendering required.

Loading

hf download greenery8848/robocost-C --repo-type dataset --local-dir ./robocost_C_ds
cd robocost_C_ds && sha256sum -c checksums.sha256 --quiet && echo ALL-MATCH

# LeRobotDataset(...) cannot load this yet (missing episodes_stats.jsonl, see known defect 0)
import pyarrow.parquet as pq, glob
t = pq.read_table(sorted(glob.glob("robocost_C_ds/data/**/*.parquet", recursive=True))[0])
print(t.column_names)          # includes cost / cost_any

License and redistribution

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

Related

Repo	Contents
robocost-A · -B · -C	packed datasets (parquet + mp4 + meta), identical schema, mergeable
robocost-raw	raw generator outputs: per-frame sim_state in 4975 traj.npz files (the only source for re-deriving costs) plus 60 verification mp4 files and _summary.log per task-level directory