NVIDIA open-sources MotionBricks for real-time character and robotics animation Other

NVIDIA open-sourced MotionBricks, a neural framework capable of generating fluid motions for video game characters and humanoid robots on the fly.

Trained on 350,000 motion-capture clips, the model synthesizes each frame in approximately two milliseconds on an NVIDIA RTX 5090 GPU, allowing game engines and control loops to evaluate movements without noticeable latency.

What it is

• Real-time motion synthesis. MotionBricks generates walking, running, vaulting over obstacles, landing, and item interactions dynamically based on velocity, target direction, and sparse keyframe poses.
• Zero-shot style variation. The model supports diverse locomotion styles ranging from injured or limping characters to zombie walks without requiring per-style retraining.
• Elimination of manual animation graphs. Technical animators no longer need to author extensive state transition trees or blend trees manually.
• Humanoid robotics control. NVIDIA demonstrated MotionBricks controlling the Unitree G1 humanoid robot in physical simulation and real-world deployment.

Performance specifications

Sources

• NVIDIA MotionBricks Project Page
• NVIDIA GEAR Lab