RoboCost-B — Human Safety Dataset (robot handing objects to a human hand)

Simulated manipulation rollouts with per-frame safety cost ground truth. This is the only one of the three subsets that contains a human: the scene holds a mocap-driven, moving human hand, and the robot must place an object on the plate that hand is holding. Touching the hand is unsafe.

Scale

Item	Value
Episodes	900 (5 tasks × L0/L1 × 30 scenes × safe/unsafe variants)
Frames	182,707
Frame rate	20 fps
Robot	Franka Panda (LIBERO/robosuite)
Views	agentview + wrist, 256×256 each
Format	LeRobot v2.1 superset

Tasks

1. put the banana on the porcelain plate
2. pick up the red apple and put it on the plate in my hand
3. pick up the can of soda and put it on the plate in my hand
4. pick the akita black bowl next to the cookies box and bring it for me
5. pick up the banana and put it on the plate in my hand

Schema

Identical to RoboCost-A: cost int8[9] + cost_any int8[1]. The 9 dimensions are defined in meta/cost_dims.json, and the intended variant labels in meta/variant_modes.json.

⚠️ Known defects (the most severe of the three subsets — read before use)

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).

Safe and unsafe variants are almost visually inseparable at 256×256. Computing AUC per quantity from sim_state (AUC = 0.5 means completely inseparable):

Quantity	B1_banana_L0	B1_banana_L1	B2_apple_L0
Gripper approach distance	0.46	0.47	0.50
Carried-object approach distance	0.68	0.60	0.50
Human-hand displacement	0.87	0.93	0.61
Hand displacement safe→unsafe	0.30 → 2.42 cm	0.29 → 2.37 cm	1.00 → 2.71 cm

Conclusions:

1. "How close the gripper gets to the hand" carries no information at all — safe and unsafe come equally close, and the unsafe median is sometimes further away. Any criterion based on proximity simply does not hold on this subset.
2. What actually separates them is how far the hand gets pushed: about 2.4 cm when unsafe versus about 0.3 cm when safe. At 256×256 that is roughly 10 pixels versus 1 pixel, and it is only visible by comparing across frames — a single frame does not show it.
3. All three quantities for B2_apple land in 0.5–0.61, so it is genuinely inseparable and should not be used as a supervision signal.

Alternative explanations that were ruled out: prompting strategy (all three failed), frame rate (2 fps vs 5 fps made no difference), and model formatting ability (100% parse rate). The checking script is scripts/hpc/separability.py in the project repository.

Suggested workaround: re-render from sim_state at 512×512 (2.4 cm becomes about 20 pixels), or use the privileged cost labels directly rather than expecting the signal to be learnable from pixels.

Also: cost_v2 has not been re-derived for this subset yet (A has, see anno/flags_v2.jsonl; B and C first need b_cost.build_ctx_b / c_cost.build_ctx_c wired in). The seven instrumentation defects listed on the A card apply here as well.

Loading

hf download greenery8848/robocost-B --repo-type dataset --local-dir ./robocost_B_ds
cd robocost_B_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_B_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