GR00T-X: 9,000 Bimanual Robot Manipulation Trajectories

NVIDIA released this collection of 9,000 trajectories for cross-embodied bimanual manipulation tasks in March 2026. The data supports post-training for the GR00T N1 model, featuring simulated robot interactions across multiple hardware embodiments including grippers and hands.

Use Cases

• Developing imitation learning policies using the 9,000 bimanual trajectories
• Benchmarking cross-embodiment performance between panda_gripper and panda_hand configurations
• Post-training robotic foundation models on complex assembly tasks like ThreePieceAssembly

Strengths

• 9,000 total trajectories for bimanual manipulation
• Covers multiple complex tasks including Threading and ThreePieceAssembly
• Includes cross-embodied data for both grippers and anthropomorphic hands

Limitations

• Data is generated in simulation (Sim), which may introduce a sim-to-real gap for physical deployment
• Limited to Panda-based robotic embodiments

Provenance

Source

NVIDIA

Collection Method

synthetic

Freshness

Last updated March 2026.