Researchers led by Zexun Lin have built WAMotor, a method that speeds up world action models, which predict both the scene and a robot's actions, according to a paper on arXiv. Such models are slow because they produce actions over many denoising steps.

WAMotor drafts and checks steps in parallel. It avoids reusing earlier velocities at the first and last steps, where they change sharply, and scales its acceptance thresholds to the length of each time step. On the LIBERO, RoboTwin 2.0 and LIBERO-Plus benchmarks it made models 2.86 to 4.28 times faster from start to finish while keeping their task success rates.

Source: arXiv Robotics (cs.RO)

Background

Zexun Lin

Robotics researcher

Robotics researcher.