GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories in VLA Policies
GeoAAC adaptively sizes action chunking horizons in flow-based VLA policies using denoising trajectory geometry, raising real-world manipulation success from 53.3% to 74.4%.
GeoAAC exploits geometric variation across action prefixes in flow-matching denoising trajectories as a process-level signal of prediction reliability. It constructs a horizon-wise geometric profile and adaptively determines the action horizon from a single generation without additional training. Experiments with GR00T N1.5 and π0.5 on LIBERO, LIBERO-Pro, and RoboCasa365 show gains up to 8.7 percentage points over fixed-horizon and adaptive baselines, with real-world success rate rising from 53.3% to 74.4%.
- Uses flow-matching trajectory geometry to gauge prediction reliability
- Adapts action horizon per rollout stage without extra training
- Validated on GR00T N1.5 and π0.5 across LIBERO and RoboCasa365
- Real-world manipulation success rises from 53.3% to 74.4%
Full article175 words · extracted from arxiv.org · click to collapse
Action chunking is widely used for action generation and execution in Vision-Language-Action (VLA) policies, yet existing approaches commonly use a fixed action horizon. During a rollout, different task stages may require different levels of action continuity, control precision, and closed-loop feedback, making a fixed horizon unable to accommodate changing control requirements. We propose \textbf{GeoAAC}, a geometry-based adaptive action chunking method for flow-based VLA policies that adjusts the action horizon according to the reliability of the current action prediction. We show that the geometry of Flow Matching denoising trajectories provides process-level information for characterizing prediction reliability, with geometric variation across action prefixes remaining positively correlated with predictive uncertainty. GeoAAC uses this prefix-wise geometry to construct a horizon-wise geometric profile and adaptively determine the action horizon from a single generation without additional training. Experiments with GR00T N1.5 and π0.5 on LIBERO, LIBERO-Pro, RoboCasa365, and real-world manipulation tasks show consistent improvements over fixed-action-horizon baselines and existing adaptive methods, including up to 8.7 percentage points in simulation and an increase in average real-world success rate from 53.3\% to 74.4\%.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.20776