GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies
GeoAAC changes a robot policy’s action horizon on the fly using geometry from its own denoising trajectory.
The paper targets VLA policies that still execute fixed-length action chunks even when a task shifts between coarse motion and precise control. GeoAAC reads prefix-wise geometric variation in Flow Matching denoising trajectories as a reliability signal, then chooses the action horizon from a single generation without extra training. In tests with GR00T N1.5 and π0.5, the authors report gains over fixed-horizon and adaptive baselines, including up to 8.7 points in simulation and real-world success rising from 53.3% to 74.4%. ArXiv · AI/CL/LG's note
The paper targets VLA policies that still execute fixed-length action chunks even when a task shifts between coarse motion and precise control. GeoAAC reads prefix-wise geometric variation in Flow Matching denoising trajectories as a reliability signal, then chooses the action horizon from a single generation without extra training. In tests with GR00T N1.5 and π0.5, the authors report gains over fixed-horizon and adaptive baselines, including up to 8.7 points in simulation and real-world success rising from 53.3% to 74.4%. ArXiv · AI/CL/LG's note
score 4