HumanoidBench Balance Simple

LeRobot v2 image-parquet exports for the g1_sonic_balance_simple task in HumanoidBench. The task prompt is:

Keep the G1 humanoid balanced on the support board.

This dataset is published in the latency-sensitive-bench/humanoidbench dataset repository. HumanoidBench scene assets are not duplicated here; use latency-sensitive-bench/g1-sonic-assets at revision 7c0a6616c78106adf9fa54dfc40e56ea07dc7d5d.

Dataset contents

The repository layout is:

balance_simple/
├── lerobot/       # train: 90 episodes, 86,048 frames
└── lerobot__val/  # validation: 10 episodes, 9,736 frames

Each split is a self-contained LeRobot v2 dataset with data/, videos/, and meta/. Data is sampled at 50 Hz; one row is one decision step. The episodes contain 100 total teacher rollouts split 90/10 with split seed 0.

Schema

Field	Type and shape	Meaning
observation.image	video, 256 × 256 × 3, H.264, 50 Hz	head-camera RGB observation
observation.state	float32[198]	native G1 teacher observation
action	float32[43]	absolute native joint target
timestamp	float64	seconds from episode start
episode_index	int64	episode id
frame_index	int64	frame id within episode
task_index	int64	index into meta/tasks.jsonl
latency	int64	recorded latency metadata
latency_raw_frames	int64	recorded latency in raw frames
done	bool	episode termination flag
reward	float32	teacher environment reward
density_weight	float32	rollout density weight

The 43 action entries are ordered as follows:

left_hip_pitch_joint, right_hip_pitch_joint, waist_yaw_joint,
left_hip_roll_joint, right_hip_roll_joint, waist_roll_joint,
left_hip_yaw_joint, right_hip_yaw_joint, waist_pitch_joint,
left_knee_joint, right_knee_joint, left_shoulder_pitch_joint,
right_shoulder_pitch_joint, left_ankle_pitch_joint, right_ankle_pitch_joint,
left_shoulder_roll_joint, right_shoulder_roll_joint, left_ankle_roll_joint,
right_ankle_roll_joint, left_shoulder_yaw_joint, right_shoulder_yaw_joint,
left_elbow_joint, right_elbow_joint, left_wrist_roll_joint,
right_wrist_roll_joint, left_wrist_pitch_joint, right_wrist_pitch_joint,
left_wrist_yaw_joint, right_wrist_yaw_joint, left_hand_thumb_0_joint,
left_hand_thumb_1_joint, left_hand_thumb_2_joint, left_hand_middle_0_joint,
left_hand_middle_1_joint, left_hand_index_0_joint, left_hand_index_1_joint,
right_hand_thumb_0_joint, right_hand_thumb_1_joint, right_hand_thumb_2_joint,
right_hand_index_0_joint, right_hand_index_1_joint, right_hand_middle_0_joint,
right_hand_middle_1_joint

State entries are ordered state_0 through state_197; the authoritative feature metadata is in each split's meta/info.json and meta/modality.json.

Provenance

The data is generated from the G1 SONIC Balance Simple teacher configured in configs/teachers/g1_sonic_gr00t.json, using scripts/teachers/export_g1_sonic_gr00t.py with --balance-teacher-observation-state, deterministic rollout, 100 maximum episodes, validation fraction 0.1, and split seed 0. The converted output is intended for NVIDIA Isaac-GR00T NEW_EMBODIMENT training.

The state is the teacher's 198D pre-decision observation. Actions are 43D absolute native joint targets. No normalization or action-dimension padding is applied in this dataset.

Intended use

Use balance_simple/lerobot for training and keep balance_simple/lerobot__val for validation. The official GR00T modality maps these fields to state.state, action.action, and video.ego_view; the 40-step action horizon uses the current row as input and repeats its current action across all 40 labels.

This dataset is for research and benchmarking in the HumanoidBench Balance Simple environment. It does not include model checkpoints or W&B artifacts.

No license metadata is asserted here.