jellyho/hmg-g1-drill-pnp-reward

HumanoidMimicGen G1 Drill PnP: success + failure episodes for reward modeling

Pick a drill from one table and place it on a second table. (HMG environment LMDrillPnP90).

1095 episodes (564 success / 531 failure), 1,269,129 frames at 50 Hz, each with a 480x640 ego-view RGB video, robot state, actions and per-frame task-success labels. Format: LeRobot v2.1 (data/ parquet + videos/ mp4 + meta/).

How it was made

Source: the public HumanoidMimicGen G1 benchmark (datasets/06_drill_pnp, ~1K generated successful demos with per-frame MuJoCo states), simulated with the HumanoidMimicGen environments (MuJoCo 3.2.6, RoboSuite 1.5.1) and the GR00T whole-body controller (WBC).

• Success episodes (clean): every recorded MuJoCo state of a benchmark demo is restored after resetting the env with the demo's seed, and the ego camera is rendered. These are the original demonstrations, re-rendered.
• Perturbed episodes: the demo is replayed exactly (recorded states) up to a perturbation step t0, then continues as closed-loop physics: the recorded upper-body / navigation goals are fed (with the perturbation applied) to the WBC, which keeps the robot balanced. The episode runs for the demo's length. Open-loop replay of the recorded joint actions is chaotic for these demos, which is why only the prefix is state-restored.
• Labels come from the environment's own success predicate (env.is_success()["task"]), evaluated every step (next.success). An episode is a success iff the predicate fired at any step. Perturbed episodes that still succeeded are kept as successes, so "physics after a branch point" is not a shortcut for failure.
• Source demos whose exact state replay never satisfied the predicate were dropped (ambiguous).
• Failures were subsampled to at most a 1:1 ratio with successes.

perturbation	what it does	success	failure
clean	exact replay of the recorded benchmark demo (all recorded MuJoCo states restored)	399	0
arm_noise	temporally correlated (OU) noise on arm joint goals	1	110
drop	hands opened while carrying the object	36	109
early_release	hands opened shortly before the place pose	84	26
miss_grasp	constant random arm-joint offset applied before/through the grasp	34	95
nav_drift	constant bias added to the base velocity command while walking	7	88
no_grasp	hands kept open for the whole episode	0	34
stall	all goals frozen and base stopped from a random step on	3	69

Features

key	shape	description
observation.images.ego_view	480x640x3 video	head (OAK) ego camera
observation.state	43	whole-body joint positions (legs, waist, arms, hands)
action	35	WBC goal: 31 upper-body joint targets + navigate (vx, vy, yaw rate) + base height (same contract as the HMG policy example)
action.joint	43	low-level joint position command actually sent
next.success	1	task predicate after this step
next.reward	1	float(next.success)
next.done	1	last step of the episode
observation.sim.mujoco_state	800	full MuJoCo state, zero padded (observation.sim.mujoco_state_len valid entries)

meta/episode_labels.jsonl has one row per episode: success, final_success, first_success_step, perturbation, its params (t0, detected grasp/release steps, magnitudes), source shard/episode and seed. Episodes from the same source demo share source_shard/source_episode; split by source to avoid leakage.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset  # lerobot 0.3.x reads v2.1
ds = LeRobotDataset("jellyho/hmg-g1-drill-pnp-reward")

License and attribution

Derived from linkenv/humanoidmimicgen-g1-benchmark (Apache-2.0) and rendered with the HumanoidMimicGen assets; see the HumanoidMimicGen repository for the licenses of RoboSuite, RoboCasa-derived assets and Unitree G1 models. If you use this data, please cite HumanoidMimicGen:

@article{lin2026humanoidmimicgen,
  title   = {HumanoidMimicGen: Data Generation for Loco-Manipulation via Whole-Body Planning},
  author  = {Kevin Lin and Ajay Mandlekar and Caelan Reed Garrett and Nikita Chernyadev and Yu Fang and
             Runyu Ding and Yuqi Xie and Justin Tran and Linxi Fan and Yuke Zhu},
  year    = {2026},
  eprint  = {2605.27724},
  archivePrefix = {arXiv}
}