cloudwalk-research/psi0-apple-to-plate-teleop

Psi0 Apple-to-Plate VR Teleoperation Dataset

Human-demonstrated loco-manipulation trajectories for fine-tuning the Psi0 Vision-Language-Action (VLA) model on a Unitree G1 humanoid robot in Isaac Lab simulation.

Dataset Summary

Property	Value
Robot	Unitree G1 29-DOF (dexterous hands + wholebody locomotion)
Task	Pick up the apple, walk left and place the apple on the plate
Episodes	79 (filtered from 81; removed ep49=22 frames < chunk_size, ep80=metadata artifact)
Total frames	86,855
Mean episode length	1,099 frames (36.6s)
Episode range	19.4s – 73.5s
FPS	30 Hz
Total duration	~48.3 min
Format	LeRobot v2.1
Simulator	NVIDIA Isaac Lab (Isaac Sim 4.5 + PhysX 5)
Scene	apple_to_plate (ported from RoboCasa/MuJoCo, RTX rendering)
Collection method	VR teleoperation (Pico 4 Ultra headset + WebXR controllers)
Locomotion	AMO policy (System-0, RL-based, 50 Hz control rate)
Size on disk	~390 MB

Task Description

The operator wears a Pico 4 Ultra VR headset and controls the G1 robot in Isaac Lab via WebXR controllers. The task is a multi-phase loco-manipulation sequence:

1. Walk forward toward the apple on the table
2. Reach and grasp the apple with the right hand
3. Walk left toward the plate
4. Place the apple on the plate

Locomotion (forward/lateral velocity, yaw rate) is commanded via the left joystick and executed by the AMO locomotion policy. Upper-body control (arms, hands, torso) is mapped from VR controller poses via analytical IK. WebRTC H.264 video feedback from the robot's egocentric camera streams back to the VR headset at 30 FPS.

Data Format

Features

Feature	Type	Shape	Description
observation.images.egocentric	video (H.264)	(480, 640, 3)	RGB egocentric camera (D435i mount, 47.6° pitch down)
states	float32	(28,)	Joint positions: hand(14) + arm(14)
action	float32	(36,)	Action targets: hand(14) + arm(14) + torso(3) + height(1) + loco(4)
timestamp	float32	(1,)	Time within episode (seconds)
frame_index	int64	(1,)	Frame index within episode
episode_index	int64	(1,)	Episode index
index	int64	(1,)	Global frame index
task_index	int64	(1,)	Task index (0 for all frames)
next.done	bool	(1,)	True on last frame of episode
obs_timestamp	float32	(1,)	Observation wall-clock timestamp
action_timestamp	float32	(1,)	Action wall-clock timestamp

Action Space (36-dim)

Index   Name              Description
─────   ────              ───────────
[0:7]   Left hand         thumb_0/1/2, index_0/1, middle_0/1
[7:14]  Right hand        thumb_0/1/2, index_0/1, middle_0/1
[14:21] Left arm          shoulder_pitch/roll/yaw, elbow, wrist_roll/pitch/yaw
[21:28] Right arm         shoulder_pitch/roll/yaw, elbow, wrist_roll/pitch/yaw
[28:31] Torso             roll, pitch, yaw
[31]    Base height       target standing height (m)
[32:35] Locomotion        vx (forward), vy (lateral), vyaw (yaw rate)
[35]    Target yaw        heading angle (rad)

State Space (28-dim)

Index   Name              Description
─────   ────              ───────────
[0:7]   Left hand         joint positions (rad)
[7:14]  Right hand        joint positions (rad)
[14:21] Left arm          joint positions (rad)
[21:28] Right arm         joint positions (rad)

Action Statistics (per joint)

Joint	Min	Max	Mean	Std
L_thumb_0	0.000	1.050	0.011	0.104
L_thumb_1	0.000	0.920	0.010	0.091
L_thumb_2	0.000	1.750	0.018	0.174
L_index_0	-1.570	0.000	-0.016	0.156
L_index_1	-1.750	0.000	-0.018	0.174
L_middle_0	-1.570	0.000	-0.016	0.156
L_middle_1	-1.750	0.000	-0.018	0.174
R_thumb_0	-1.050	0.000	-0.364	0.410
R_thumb_1	-0.920	0.000	-0.319	0.360
R_thumb_2	-1.750	0.000	-0.607	0.684
R_index_0	0.000	1.570	0.544	0.614
R_index_1	0.000	1.750	0.607	0.684
R_middle_0	0.000	1.570	0.544	0.614
R_middle_1	0.000	1.750	0.607	0.684
L_shoulder_p	-0.553	0.379	-0.159	0.162
L_shoulder_r	-0.201	0.584	0.268	0.105
L_shoulder_y	-0.064	0.604	0.246	0.095
L_elbow	-0.740	1.367	1.126	0.466
L_wrist_r	-1.012	0.611	-0.065	0.117
L_wrist_p	-1.028	0.146	-0.396	0.176
L_wrist_y	-0.621	0.348	-0.007	0.108
R_shoulder_p	-1.648	0.148	-0.485	0.300
R_shoulder_r	-0.726	0.323	-0.068	0.177
R_shoulder_y	-0.943	0.537	-0.153	0.234
R_elbow	-0.586	1.362	0.368	0.474
R_wrist_r	-0.700	1.681	0.351	0.522
R_wrist_p	-1.036	0.571	-0.249	0.329
R_wrist_y	-0.514	1.325	0.334	0.343
torso_roll	0.000	0.000	0.000	0.000
torso_pitch	0.391	0.727	0.540	0.058
torso_yaw	-0.363	0.317	-0.005	0.075
base_height	0.750	0.750	0.750	0.000
vx	-0.759	0.757	0.011	0.345
vy	-0.757	0.760	0.047	0.262
vyaw	-0.750	0.750	-0.001	0.252
target_yaw	0.000	0.000	0.000	0.000

Collection Pipeline

Pico 4 Ultra (VR)  ──WebXR──►  teleop_bridge.py (IK)  ──ZMQ──►  Isaac Lab (PhysX 500Hz)
       ▲                                                              │
       └──────────── WebRTC H.264 30fps ◄── TiledCamera 640x480 ─────┘
                                                                      │
                                                              data_recorder.py
                                                              (LeRobot format)

• Simulator: NVIDIA Isaac Lab 2.3.2 (Isaac Sim 4.5, PhysX 5, RTX rendering)
• Physics: 500 Hz internal, 50 Hz control rate (AMO substeps=10)
• Camera: TiledCamera 640x480 RGB mounted on head_link, matching RealSense D435i position and angle (47.6° pitch down)
• Locomotion: AMO policy (System-0) — RL-based, ported from IsaacGym. Inputs: 1043+2325 dim proprioception. Outputs: 15-dim leg PD targets
• Action semantics: Absolute joint positions (not deltas)
• Normalization: bounds type (see meta/stats_psi0.json)
• Action-observation sync: p50 = 50ms, p95 = 66ms

Directory Structure

├── data/chunk-000/
│   ├── episode_000000.parquet
│   ├── episode_000001.parquet
│   └── ... (79 files)
├── videos/chunk-000/egocentric/
│   ├── episode_000000.mp4
│   ├── episode_000001.mp4
│   └── ... (79 files)
├── meta/
│   ├── info.json
│   ├── episodes.jsonl
│   ├── tasks.jsonl
│   ├── episodes_stats.jsonl
│   └── stats_psi0.json
└── README.md

Intended Use

This dataset is designed for fine-tuning the Psi0 VLA model using finetune_simple_psi0_config with SimpleRepackTransform and bounds normalization. The training pipeline expects:

• Action chunk size: 30 frames
• VLM backbone: frozen (Qwen3-VL-2B-Instruct)
• Action expert: MM-DiT (~500M params, flow matching)
• Batch size: 128 (16/GPU x 8 GPUs)
• Training steps: 40,000

Data Cleaning

This dataset was filtered from 81 to 79 episodes using scripts/data/filter_episodes.py:

Removed	Reason
Episode 49 (original)	Only 22 frames (0.7s) — shorter than action_chunk_size=30, would be padded during training
Episode 80 (original)	Metadata artifact from info.json off-by-one in original recording pipeline

Remaining episodes were renumbered contiguously (0–78). All global indices, frame indices, and normalization stats were recomputed. Validated with scripts/data/validate_dataset_preflight.py --strict.

Citation

If you use this dataset, please cite both this dataset and the original Ψ₀ project:

@misc{setubal2026psi0appletoplate,
  title={Psi0 Apple-to-Plate VR Teleoperation Dataset},
  author={Pedro Setubal and CloudWalk Research Lab},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/cloudwalk-research/psi0-apple-to-plate-teleop}},
}

This dataset is built for fine-tuning the Ψ₀ VLA model. Please also cite the original work:

@misc{wei2026psi0,
  title={$\Psi_0$: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation},
  author={Songlin Wei and Hongyi Jing and Boqian Li and Zhenyu Zhao and Jiageng Mao and Zhenhao Ni and Sicheng He and Jie Liu and Xiawei Liu and Kaidi Kang and Sheng Zang and Weiduo Yuan and Marco Pavone and Di Huang and Yue Wang},
  year={2026},
  eprint={2603.12263},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2603.12263},
}