Unitree G1 Robot Teleoperation with Forward Kinematics Augmentation

Augmented version of NVIDIA's PhysicalAI-Robotics-GR00T-Teleop-G1 release, adding forward-kinematics signals for the robot's palms. It is designed to facilitate imitation learning and planning for an Unitree G1 humanoid robot performing fruit pick-and-place tasks. The dataset was created by Nirav-Madhani and was last updated in September 2025.

Use Cases

• Train an imitation learning policy using the left_palm_pose and right_palm_pose signals as compact goal representations.
• Develop a model predictive control (MPC) system leveraging the short-horizon future traces of palm poses for planning.
• Analyze teleoperation trajectories for fruit pick-and-place tasks using the forward-kinematics features.

Strengths

• Augments a foundational robotics teleoperation dataset from NVIDIA with additional forward-kinematics signals.
• Provides goal-centric features like left_palm_pose and right_palm_pose designed to simplify downstream learning tasks.
• Dataset was updated in September 2025, indicating recent maintenance.

Limitations

• The dataset size, row count, and specific column definitions are unknown, limiting assessment of scale and structure.
• The raw description is a summary; full details on data collection, sensor specifications, and potential biases are not provided here.
• The dataset's utility is tied to the original NVIDIA teleoperation data, which may have its own limitations not described.

Provenance

Source

NVIDIA's PhysicalAI-Robotics-GR00T-Teleop-G1 release, augmented by Nirav-Madhani.

Collection Method

Teleoperation of a Unitree G1 humanoid robot, with added forward-kinematics computations for palm poses.

Freshness

Last updated on 2025-09-21.