BATON makes robot exploration additive

BATON, a new robot-manipulation agent, reports an 11.6% task-success improvement and a 14.9% cumulative-success improvement on RoboMemArena by exploring subtasks separately instead of relearning the whole chain. The paper’s transition-aware memory design stores entry and exit conditions so one skill leaves the scene usable by the next; the authors argue that this changes exploration cost from roughly T^K for K stages to T×K. For builders, the useful idea is not another VLA checkpoint but a memory contract that records whether each handoff is executable.

The result belongs beside the site’s analysis of the reliability gap in business agents: long workflows fail at the seams, not only inside individual steps.