Don't Drop the BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory
BATON turns long robot tasks into separately explored subtasks, then stitches them with transition-aware memory.
The paper argues that whole-task test-time exploration scales poorly because each added stage multiplies the search cost. BATON instead explores each subtask in a short-horizon setting, stores the result, and composes the full trajectory from those pieces. It also checks whether the scene is ready before invoking the VLA model, restores usable handoff states between subtasks, and chooses strategies with the next step in mind. The authors report gains of 11.6% in task success and 14.9% in cumulative success on RoboMemArena, without updating model parameters. ArXiv · AI/CL/LG's note
The paper argues that whole-task test-time exploration scales poorly because each added stage multiplies the search cost. BATON instead explores each subtask in a short-horizon setting, stores the result, and composes the full trajectory from those pieces. It also checks whether the scene is ready before invoking the VLA model, restores usable handoff states between subtasks, and chooses strategies with the next step in mind. The authors report gains of 11.6% in task success and 14.9% in cumulative success on RoboMemArena, without updating model parameters. ArXiv · AI/CL/LG's note
score 5