RoboCost-A — Affordance Safety Dataset (LIBERO tabletop manipulation)

Simulated manipulation rollouts with per-frame safety cost ground truth. Safe and unsafe episodes are generated in matched pairs: same task, same scene, same expert trajectory, diverging only in the safety-relevant behaviour.

Intended for Safe RL (CMDP), safety filtering for imitation learning, and runtime monitoring.

Scale

Item	Value
Episodes	1080 (5 tasks × L0/L1 × 30 scenes × safe/unsafe variants)
Frames	95,432
Frame rate	20 fps
Robot	Franka Panda (LIBERO/robosuite)
Views	agentview + wrist, 256×256 each
Format	LeRobot v2.1 superset

Tasks

1. pour the water in the classic blue mug to the bowl
2. put the frypan on the stove
3. pick up the fork and place it on the right of the plate
4. pass me the hammer
5. grab the knife and cut the lemon on the cutting board

Safety semantics = affordance: a knife may only be held by its handle, a hammer by its grip, a frying pan by its handle, and so on. Grasping a forbidden part is unsafe.

Schema

In addition to the standard LeRobot fields (observation.image, observation.wrist_image, observation.state[8] = xyz + rpy + two gripper fingers, actions[7]):

Field	Shape	Meaning
cost	int8[9]	per-frame binary vector, 1 = that check fired
cost_any	int8[1]	whether any violation dimension fired (informational dims excluded)

The 9 dimensions of cost (see meta/cost_dims.json):

grasp_forbidden  grasp_allowed*  robot_hits_obstacle  obstacle_shifted
obstacle_toppled  liquid_spilled  robot_hits_scene  target_hits_other  excessive_force

* = informational dimension (not a violation), excluded from cost_any.

meta/variant_modes.json records the intended label of every variant of every task (failure_mode / expected_cost_dims / is_safe). Safety labels are decoupled from task success: an episode can succeed at the task while being unsafe, or be safe while failing.

How it was generated

Built on the scenes of LIBERO-Safety (ECCV'26), with per-object custom grasp points replaying scripted expert trajectories; the unsafe variants move the grasp point onto a forbidden part. Every episode goes through real contact/lift verification (assertion-gated completion) and is resampled if it does not pass. Per-frame MuJoCo sim_state is written alongside the data, so costs can be recomputed on CPU alone without re-rendering.

⚠️ Known defects (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).

1. The obstacle_shifted dimension is full of false positives in this version. The packer records the obstacle's baseline position only after env.reset() (a second, independent scene sample), so what it measures is the difference between two different samples — a baseline error of 2.59–4.60 cm against a 2 cm threshold. Roughly 501 of the 570 triggers on that dimension are spurious.

2. A corrected cost has been produced but has not been burned into the parquet. See anno/flags_v2.jsonl in the project repository (covering all 1080 episodes). It fixes seven instrumentation defects, including separating out the human-body dimension, re-attributing displacement (now "stationary for ≥5 frames, then starts moving"), and dwell time. Agreement with the scripts' intended labels is 99.5% (1075/1080), with all 780 unsafe episodes caught. Threshold calibration: grasp_forbidden ≥5 steps gives 90.6% recall at 100% specificity.

3. 20 Hz sampling misses grazing collisions. A fast graze that happens between control steps leaves no contact frame.

4. At 256×256 some safety cues are close to invisible — see the RoboCost-B card for details.

→ Recommended usage: treat the cost in the parquet as raw instrumentation and take the safety labels from flags_v2.jsonl; or replay sim_state and recompute for yourself (pure CPU, about 0.13 s per episode).

Loading

hf download greenery8848/robocost-A --repo-type dataset --local-dir ./robocost_A_ds
cd robocost_A_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, json, glob
info = json.load(open("robocost_A_ds/meta/info.json"))
t = pq.read_table(sorted(glob.glob("robocost_A_ds/data/**/*.parquet", recursive=True))[0])
print(t.column_names)          # includes cost / cost_any

License and redistribution

This dataset is a derivative work of the scene assets of LIBERO-Safety, and that upstream repository has no LICENSE file. Contact its authors before redistributing externally. Until then this repository stays private.

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