npow/yam-gello-red-cap-100

YAM GELLO Red Cap Manipulation Dataset (100 Demonstrations)

This dataset contains 100 high-quality bimanual teleoperation demonstrations of a tabletop object grasping and manipulation task using the I2RT YAM robot arm teleoperated with a GELLO passive leader arm.

---

Dataset Overview

• Robot: I2RT YAM manipulator (6-DOF + parallel jaw gripper, DM-series actuators over SocketCAN).
• Teleoperation Interface: GELLO leader arm (Dynamixel XL330-M077 servos, 57,600 baud).
• Task: Reaching, grasping, and manipulating a red cylindrical bottle cap / container on a tabletop.
• Total Demonstrations: 100 episodes (episode_0001 through episode_0100).
• Total Frames: 25,874 timesteps (~20-30 Hz control loop).
• Total Active Duration: ~22 minutes of continuous teleoperation.
• Camera Observations: 3 synchronized RGB cameras (640x480 H.264 @ 15-30 fps):
  • left.mp4: Left arm wrist camera (egocentric perspective)
  • middle.mp4: Overhead workspace camera (allocentric perspective)
  • right.mp4: Right arm wrist camera
• Proprioception & Action Data (trajectory.npz):
  • timestamp: High-precision float64 epoch timestamps
  • follower_joints: Real-time joint positions from CAN bus (14 DOFs: 6 left + 1 gripper, 6 right + 1 gripper)
  • leader_actions: Commanded joint targets from the GELLO leader
  • tracking_error: Absolute tracking error between leader and follower
• Interactive Visualization: Native recording.rrd streams included per episode for instant inspection in the Rerun Viewer.

---

Quality Metrics & Statistics

All 100 episodes passed automated data quality validation:

• Episode Duration: Min 9.6s, Median 12.4s, Max 24.3s (0 abandoned or accidental rollouts).
• Tracking Error: Average mean tracking error = 0.004 rad (0.2°). Zero joint tracking dropouts.
• Gripper Actuation: 100% (100/100) of episodes feature full open-and-close grasp cycles.
• Camera Sync: 300/300 video streams decoded with 1:1 frame synchronization to trajectory timesteps.

---

Directory Structure

data/episodes/
├── episode_0001/
│   ├── left.mp4            # Left wrist camera (640x480)
│   ├── middle.mp4          # Overhead workspace camera (640x480)
│   ├── right.mp4           # Right wrist camera (640x480)
│   ├── trajectory.npz      # Joint states, leader actions, timestamps
│   ├── metadata.json       # Episode metadata (num_frames, date)
│   └── recording.rrd       # Rerun native recording stream
├── ...
└── episode_0100/

---

Quick Start & Usage

1. Download Dataset with huggingface_hub

huggingface-cli download npow/yam-gello-red-cap-100 --repo-type dataset --local-dir ./yam-red-cap

2. Loading Trajectory Data in Python

import numpy as np

data = np.load("data/episodes/episode_0100/trajectory.npz")
print("Keys:", list(data.keys()))
print("Timesteps:", len(data["timestamp"]))
print("Follower Joints shape:", data["follower_joints"].shape) # (N, 14)
print("Leader Actions shape:", data["leader_actions"].shape)   # (N, 14)

3. Interactive Visualization with Rerun

pip install rerun-sdk
rerun data/episodes/episode_0100/recording.rrd

---

License & Citation

• Software based on GELLO.
• Hardware based on I2RT YAM and GELLO Mechanical.
• Dataset released under the MIT License.