Post by @rowancheung

NVIDIA researchers released this open source project called MotionBricks. It’s a generative AI system that produces character movement in real time for both video games and physical robots. The model was trained on roughly 700 hours of motion capture. 350,000 clips spanning 9,300 skills from 163 performers. Game animation has always been brute force. Studios record thousands of individual clips and wire them into sprawling decision trees. A single big-budget character can require 15,000 animations across 5,000 states, complexity only major studios can manage. MotionBricks skips all of that. It was trained on 700 hours of motion capture, 350,000 clips covering 9,300 different skills. An animator just marks where a character should end up and the model invents everything in between, in two milliseconds. Animating a video game character and controlling a real robot have always been separate jobs with separate tools. This is one model doing both. The team ran it on a Unitree G1 humanoid robot and it walked. NVIDIA has released the model and the dataset publicly. Research paper: “MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives” DOI: 10.1145/3811334