GeoAAC: Geometry-Based Adaptive Action Chunking from Denoising Trajectories for Flow-Based VLA Policies

Forum topic · 小凯 · 2026-09-19

Summary

GeoAAC is a geometry-based adaptive action chunking method for flow-based Vision-Language-Action (VLA) policies, introduced in arXiv paper 2609.20776 by Xin Chen and colleagues. Unlike existing approaches that use a fixed action horizon, GeoAAC adjusts the horizon according to the reliability of the current action prediction. The key insight is that the geometry of Flow Matching denoising trajectories provides process-level signals for prediction reliability: geometric variation across action prefixes remains positively correlated with prediction uncertainty. By building horizon-level geometry profiles from prefix-level geometry, GeoAAC adaptively determines the action horizon from a single generation pass without additional training. Experiments on GR00T N1.5 and π0.5 across LIBERO, LIBERO-Pro, RoboCasa365, and real-world manipulation tasks consistently outperform fixed-horizon baselines and prior adaptive methods, with gains up to 8.7 percentage points in simulation and real-world average success rates improving from 53.3% to 74.4%.

Paper Overview

Field: Machine Learning (ML) Authors: Xin Chen, Sen Chen, Yujuan Ding, Jian Liu, Guoqing Wang, Wei Ye, Heng Tao Shen, Yi Bin Published: 2026-09-17 arXiv: 2609.20776

Abstract

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.

The authors propose 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. Key contributions:

The geometry of Flow Matching denoising trajectories provides process-level information for characterizing prediction reliability.
Geometric variation across action prefixes remains positively correlated with prediction uncertainty.
GeoAAC leverages this prefix-level geometry to construct horizon-level geometry profiles, adaptively determining the action horizon from a single generation without additional training.

Results

Evaluated on GR00T N1.5 and π0.5 across LIBERO, LIBERO-Pro, RoboCasa365, and real-world manipulation tasks, GeoAAC consistently outperforms fixed action horizon baselines and existing adaptive methods:

Up to 8.7 percentage points improvement in simulation
Real-world average success rate improved from 53.3% to 74.4%

--- *Auto-collected on 2026-09-19.*