BCP: Adaptive Horizon Execution for Vision-Language-Action Models
A novel framework named Bernoulli-Continuation Policy (BCP) has been introduced to overcome the challenges faced by chunk-based Vision-Language-Action (VLA) models, which generally perform a predetermined number of actions before initiating replanning. This predetermined execution horizon is uniform and periodic, not taking task advancement into account, which can result in crucial manipulation phases being carried out based on outdated chunks if no replanning occurs beforehand. BCP serves as a lightweight, adaptable framework that permits dynamic execution horizons while maintaining the stability of the base VLA. Its continuation head breaks down the selection of execution horizons into a series of decisions about whether to continue or replan, applying an ordinal, prefix-sharing inductive bias. Training for the head employs reinforcement learning since the ideal horizon for each chunk is not directly observable. The research, aimed at enhancing VLA models' performance in robotic manipulation, is available on arXiv with the identifier 2608.03483v1 and was announced as a cross-type submission.
Key facts
- BCP is a framework for adaptive horizon execution in VLA models.
- It addresses the fixed execution horizon limitation in chunk-based VLA models.
- BCP is lightweight and plug-and-play, keeping the base VLA frozen.
- It uses a continuation head to make continue-or-replan decisions.
- The head imposes an ordinal, prefix-sharing inductive bias.
- Training uses reinforcement learning because optimal horizons are not observable.
- The paper is on arXiv with ID 2608.03483v1.
- The announcement type is cross.
Entities
Institutions
- arXiv