Paper Overview
Field: Robotics Authors: Bingxin Xu, Yuzhang Shang, Emilio Ferrara Published: 2026-08-17 arXiv: 2608.16889
Summary
Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task. Vision-language-action (VLA) models increasingly master the individual skills, yet the chain still fails: errors compound beyond the policy's ability to correct, and one subtask silently constrains the next.
A promising recipe freezes the VLA and puts an LLM agent in charge: it plans in language, moves in free space with analytic primitives, invokes the VLA only for contact-rich segments, and writes adaptation into language memory. Applied to long horizons, this breaks in two ways:
1. Exploration cost: Competence comes from whole-task exploration at test time, whose cost is multiplicative in stages. If one stage needs T episodes, a K-stage task needs about T^K episodes, and a failure does not reveal which stage caused it. 2. No transition representation: VLA primitives only have exit conditions, not entry conditions, so a subtask may succeed in a form its successor cannot use.
The BATON Approach
- Subtask-level exploration: Each subtask is explored within a cheap, short horizon; solutions are stored in memory, and long-horizon trajectories are composed from these solutions. Cost becomes additive (T*K), and every failure is attributable to a single stage.
- Transition-aware memory: A call transition lets the agent control when to invoke primitives; a handoff transition restores the entry state perturbed by the predecessor; a lookahead transition selects policies whose outcomes the successor can inherit.
- No parameter updates: All adaptation is written into language memory; the VLA stays frozen.
Results
On the long-horizon benchmark RoboMemArena, BATON improves task success rate by 11.6% and cumulative success rate by 14.9% over the state of the art.
Original Abstract (excerpt)
> Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task. Vision-language-action (VLA) models increasingly master the individual skills, yet the chain still fails: errors compound beyond the policy's ability to correct, and one subtask silently constrains the next. A promising recipe freezes the VLA and puts an LLM agent in charge: it plans in language, moves in free space with analytic primitives, invokes the VLA only for contact-rich segments, and writes adaptation into language memory. Applied to long horizons, it breaks twice. (1) Competence comes from whole-task exploration at test time, whose cost is multiplicative in stages: if one stage needs T episodes, a K-stage task needs about T^K, and a failure does not reveal which stage caused it. (2) It has ...
---
*Auto-collected on 2026-08-19*