G1 Robot Stacking Cups Dataset for Imitation Learning

102 episodes of robot manipulation data collected using a Unitree G1 Brainco robot. The dataset contains 22,777 frames of video and sensor data recorded at 30 fps, created by n1robotics and released in April 2026.

Use Cases

• Train a policy to predict robot actions from video frames for the cup-stacking task.
• Benchmark imitation learning models using the 22,777 frames of state-action pairs.
• Analyze temporal consistency and task progress across the 102 recorded episodes.
• Validate sim-to-real transfer by comparing learned policies against the demonstrated trajectories.

Strengths

• 102 distinct task episodes for model training
• 22,777 total frames providing dense temporal coverage
• Multimodal data includes video files totaling 500 MB

Limitations

• Limited to a single task (stacking cups), reducing generalizability
• Moderate dataset size with 102 episodes may not capture full task variability
• Geographic and hardware bias from a single robot type in one lab

Provenance

Source

n1robotics via Hugging Face

Collection Method

Collected via teleoperation or demonstration using the LeRobot framework.

Freshness

Last updated April 2026.