vikram-avea/yam-physical-ai-hack-mega-640x480

YAM Physical AI Hack MEGA Dataset (640x480@30fps)

The ultimate LeRobot v3.0 dataset for the Physical AI Hack featuring YAM robots (LimX Sentinel humanoids).

Dataset Info

• Robot: YAM (LimX Sentinel humanoid)
• Resolution: 640x480
• FPS: 30 (TRUE 30fps, all re-encoded for consistency)
• Total Episodes: 108
• Total Frames: ~103,000+
• Task: Teleoperation demonstrations
• Format: LeRobot v3.0

What's Included

This mega dataset merges 4 high-quality recording sessions:

• Demo-06: 32 episodes (re-encoded from 31fps to 30fps)
• Demo-07: 15 episodes (native 30fps)
• Demo-08: 50 episodes (native 30fps) - Largest session
• Demo-09: 11 episodes (native 30fps)

All sessions normalized to true 640x480@30fps for perfect timestamp alignment.

Why This Dataset?

✅ Largest YAM dataset - 108 episodes of diverse teleoperation
✅ TRUE 30fps - All videos re-encoded for temporal consistency
✅ Optimal resolution - 640x480 balances quality and training speed
✅ Perfect alignment - Accurate action-observation temporal relationships
✅ Production ready - Extensively validated and tested

Usage

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
from torch.utils.data import DataLoader

# Load the mega dataset
dataset = LeRobotDataset("vikram-avea/yam-physical-ai-hack-mega-640x480")
print(f"Episodes: {len(dataset)}")  # 108
print(f"Frames: {dataset.num_frames}")  # ~103,000
print(f"FPS: {dataset.fps}")  # 30

# Training loop
dataloader = DataLoader(dataset, batch_size=32, shuffle=True, num_workers=4)

for batch in dataloader:
    # Observations
    joint_state = batch["observation.state"]  # [32, 12]
    wrist_img = batch["observation.images.robot1_wrist"]  # [32, 3, 480, 640]
    overhead_img = batch["observation.images.overhead"]  # [32, 3, 480, 640]
    
    # Actions
    action = batch["action"]  # [32, 8]
    
    # Your training code here
    loss = model(joint_state, wrist_img, overhead_img, action)
    loss.backward()

Features Schema

• observation.state: Float32[12] - Joint positions (6 arm + 6 gripper states)
• action: Float32[8] - Joint commands (6 arm + 2 gripper commands)
• observation.images.robot1_wrist: Uint8[3, 480, 640] - Wrist RGB camera
• observation.images.overhead: Uint8[3, 480, 640] - Overhead RGB camera
• timestamp: Float32[1] - Frame timestamp (episode-relative)
• episode_index: Int64[1] - Episode ID (0-107)
• frame_index: Int64[1] - Frame ID within episode

About YAM Robots

YAM robots are LimX Sentinel humanoid platforms designed for dexterous bimanual manipulation:

• Dual 6-DOF arms with parallel jaw grippers
• Real-time teleoperation via VR controllers
• Compliant actuation for safe human interaction
• Optimized for complex manipulation tasks

This dataset was collected during the Physical AI Hack 2026 event.

Data Quality

All episodes have been validated for:

• Consistent 30fps across all videos
• Proper timestamp alignment
• Complete action-observation pairs
• No corrupted frames or data gaps

Citation

If you use this dataset in your research, please cite:

@dataset{yam_physical_ai_hack_mega_2026,
  title={YAM Physical AI Hack MEGA Dataset},
  author={Avea Robotics},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/vikram-avea/yam-physical-ai-hack-mega-640x480}}
}

License

MIT License - Free for research and commercial use.

Related Datasets

Individual sessions available separately:

• sentinel-demo-06 - 32 episodes
• sentinel-demo-07 - 15 episodes
• sentinel-demo-08 - 50 episodes
• sentinel-demo-09 - 11 episodes