zhouqh/hr_v2

Genesis HR Bench VLA Demonstrations

Scripted-expert demonstrations for human-robot collaborative manipulation in the Genesis simulator. This release contains exactly 300 validated successful episodes for each of 19 tasks: 5,700 episodes and 1,517,474 policy steps in total.

The robot is a single-arm Franka. Observations include head-camera and right-wrist RGB-D streams, robot state, end-effector pose, and gripper state. Actions include absolute joint targets, joint deltas, end-effector deltas, and gripper targets. Collection uses 500 Hz simulation with a capture stride of 25, giving a 20 Hz policy rate.

Tasks

• categorize_cooperative
• categorize_interrupt
• deliver_to_human_easy
• dump_bin_assist
• dump_bin_interrupt
• place_bread_in_basket_interrupt
• place_bread_in_basket_neutral
• place_burger_fries_assist
• place_burger_fries_interrupt
• place_burger_fries_neutral
• place_dual_shoes_assist
• place_dual_shoes_interrupt
• place_food_in_skillet_assist
• pour_water
• put_object_cabinet_interrupt
• put_object_cabinet_neutral
• stack_bowls_three_assist
• stack_bowls_three_interrupt
• stack_bowls_three_neutral

Tasks are added to this directory only after 300 validated successful episodes are available.

Layout

<task>/rasterizer/seed_<N>/
├── meta.json
├── steps.h5
└── verdict.json

Each episode contains only these three regular files. There are no symlinks, failed attempts, in-flight episodes, debug logs, or surplus episodes.

Repository-level metadata:

• manifest.jsonl: one row per episode, used as the Hub viewer index
• dataset_summary.json: selection policy and per-task inventory
• task_pool.txt: task names in this release
• SHA256SUMS: checksums for every other file in the release

steps.h5 schema

The HDF5 attribute schema_version is vla_steps_v2_qpos_target; T is the number of policy steps in an episode.

Key	Shape	Dtype	Description
images/head_camera	(T,)	bytes	JPEG-encoded 480×640 RGB
images/right_wrist	(T,)	bytes	JPEG-encoded 480×640 RGB
depth/head_camera	(T,480,640)	float16	Depth in metres
depth/right_wrist	(T,480,640)	float16	Depth in metres
qpos	(T,7)	float32	Measured arm joint positions
qpos_target	(T,7)	float32	Arm PD targets
ee	(T,7)	float32	End-effector [x,y,z,qw,qx,qy,qz]
gripper_meas	(T,)	float32	Measured gripper state
gripper_target	(T,)	float32	Target gripper state
qpos_target_action	(T,7)	float32	Absolute joint-target action
qpos_delta_action	(T,7)	float32	Joint-delta action
ee_delta_action	(T,6)	float32	End-effector delta action
gripper_target_action	(T,)	float32	Absolute gripper action
gripper_action	(T,)	float32	Gripper delta/action label
t	(T,)	int64	Simulation step index

The boolean qpos_target_valid and qpos_target_action_valid arrays mark valid labels. Quaternions use (w, x, y, z) ordering.

meta.json contains the task instruction, seed, camera calibration, complete configuration snapshot, success result, and safety metrics. verdict.json contains the validation decision, success/plan status, error fields, and safety evaluation.

Loading one episode

from huggingface_hub import snapshot_download
import h5py
import io
import json
import numpy as np
from PIL import Image

root = snapshot_download(
    "zhouqh/hr_v2",
    repo_type="dataset",
    allow_patterns=["pour_water/rasterizer/seed_0/*"],
)
episode = f"{root}/pour_water/rasterizer/seed_0"

with open(f"{episode}/meta.json") as f:
    meta = json.load(f)

with h5py.File(f"{episode}/steps.h5") as h:
    qpos = h["qpos"][:]
    rgb = np.stack([
        np.asarray(Image.open(io.BytesIO(bytes(frame))))
        for frame in h["images/head_camera"]
    ])

print(meta["instruction"], qpos.shape, rgb.shape)

Use allow_patterns=["<task>/**"] to download one task without downloading the complete dataset.

Selection and provenance

Episodes were collected with the Genesis rasterizer backend and selected from validated successful rollouts. For each task, the first 300 successes in ascending numeric-seed order were retained. dataset_summary.json records the selected seed range and candidate inventory, and manifest.jsonl records each episode's task, seed, step count, cameras, schema, size, and source collection.

Limitations

• This release contains successful scripted-expert demonstrations only; it is not an unbiased sample of expert success or failure rates.
• All observations are synthetic simulator outputs. Sim-to-real transfer is not evaluated by this dataset.
• HDF5 trajectory files are the canonical data. The Hub viewer displays the episode-level manifest.jsonl index rather than decoding trajectory frames.
• Safety metrics reflect the configured simulator geometry and checks recorded during collection.

License and citation

The dataset is released under the MIT license. If you use it, please cite the Genesis HR Bench project and this dataset repository.