tysyuvraj/monster-tray-pickplace

Monster Tray Pick-and-Place (Unitree G1)

A teleoperated manipulation dataset collected on a Unitree G1 humanoid (dual-arm, Dex1 grippers, dual camera) for a pick-and-place task: pick up a monster energy drink can and place it on a serving tray. Two object variants are covered — a green can and an orange can — each with its own language-conditioned task instruction.

This dataset was collected to fine-tune GR00T N1.7 for the same task; the resulting checkpoint is published at tysyuvraj/GR00T-N1.7-monster-tray-pickplace. It is also the manipulation policy ("body") supervised by an anticipatory, Qwen3-Omni-based human-robot-interaction monitor ("brain") in an ongoing research project on real-time multimodal intent prediction for HRI.

Dataset Details

• Robot: Unitree G1 humanoid, upper-body dual-arm, Dex1 dexterous grippers, 2 cameras (head-mounted + wrist-mounted)
• Collection method: human teleoperation
• Tasks (language-conditioned, one instruction string per task):
  • pick up the green monster can and place it on the tray.
  • pick up the orange monster can and place it on the tray.
• Format: LeRobot dataset format (Parquet), single train split
• Size: 18,296 rows (frames), 1.14 GB
• Columns: observation.state, action, timestamp, frame_index, episode_index, index, task_index

Episode count: recorded as 104 episodes at collection time. Please verify the exact distinct-episode count against meta/episodes in the dataset itself before citing this number externally — it was not re-verified against the currently published version at card-writing time.

Intended Use

Training or fine-tuning vision-language-action (VLA) policies for humanoid dual-arm pick-and-place manipulation, and as a benchmark scene for HRI systems that supervise or interrupt a running manipulation policy (the motivating use case for this specific dataset).

Known Limitations

• Cans were recorded empty (no internal weight). Downstream policy evaluation showed the resulting policy can occasionally knock over an empty can during approach; this is a property of the training distribution (passive stability), not a labeling or collection error.
• Only two object instances (one green can, one orange can) and one receptacle (a single tray) are covered — this is a narrow-object-set dataset, not a general can-grasping dataset.
• Camera aspect ratios differ between the head and wrist cameras; any downstream processing pipeline that assumes a single consistent aspect ratio across camera views should account for this explicitly.

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("tysyuvraj/monster-tray-pickplace")

Citation

If you use this dataset, please cite it as:

@misc{singh2026monstertray,
  author = {Singh, Yuvraj},
  title  = {Monster Tray Pick-and-Place (Unitree G1)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tysyuvraj/monster-tray-pickplace}
}