GeoAAC Teaches Robot Policies When to Look Further Ahead

Most Vision-Language-Action (VLA) robot policies — AI systems that turn camera images and language instructions directly into robot motor commands — generate their actions in batches called 'chunks,' committing to a fixed number of future steps before checking sensors again. As reported in a new arXiv preprint, researchers argue this fixed-length approach is a poor fit for real tasks, since some moments (like a delicate grasp) demand tight, frequent feedback while others (like a long reach) can tolerate committing further ahead.

Their proposed method, called GeoAAC (Geometry-Based Adaptive Action Chunking), targets VLA policies built on Flow Matching, a generative technique that produces actions through a step-by-step 'denoising' process. The team found that the geometric pattern of this denoising trajectory — specifically, how much it varies across different prefixes of the predicted action sequence — correlates with how uncertain the prediction actually is. GeoAAC uses this signal to build what the authors call a 'horizon-wise geometric profile,' allowing the robot to pick an appropriate action horizon on the fly, from a single generation pass, with no additional model training required.

The method was evaluated using two existing VLA policies, GR00T N1.5 and π0.5, across the LIBERO, LIBERO-Pro, and RoboCasa365 simulation benchmarks, as well as real-world manipulation tasks. Compared to fixed-horizon baselines and other existing adaptive-chunking methods, GeoAAC delivered consistent gains — up to 8.7 percentage points in simulated task success — and a substantial jump in real-world performance, from an average success rate of 53.3% with fixed horizons to 74.4% with the adaptive approach.

The paper, submitted for consideration at the IEEE International Conference on Robotics and Automation (ICRA) 2027, frames the work as a training-free way to make existing flow-based VLA systems more robust to the varying precision demands of real manipulation tasks — without redesigning the underlying policy itself.