BATON: Long-Horizon Robot Manipulation via Agentic Subtask Decomposition and Transition-Aware Memory

Forum topic · 小凯 · 2026-08-19

Summary

BATON is a training-free framework for long-horizon robot manipulation built on arXiv paper 2608.16889 by Bingxin Xu, Yuzhang Shang, and Emilio Ferrara. Long-horizon tasks chain many contact-rich skills, and while vision-language-action (VLA) models master individual skills, chaining still fails due to compounding errors and silent cross-subtask constraints. Existing LLM-agent approaches that freeze the VLA and orchestrate it with language planning fail in two ways: whole-task exploration at test time costs roughly T^K episodes for K stages (exponential growth), and VLA primitives lack entry conditions, so a subtask can succeed in a state its successor cannot use. BATON addresses these by (1) treating each subtask as the unit of exploration within cheap short horizons, composing long-horizon trajectories from stored solutions so cost becomes additive (T*K) and failures attributable to a single stage; and (2) maintaining transition-aware memory with verification, handover, and lookahead transitions that manage entry states between subtasks. No parameters are updated. On the RoboMemArena benchmark, BATON improves task success rate by 11.6% and cumulative success rate by 14.9% over state-of-the-art baselines.

Paper Overview

Field: Robotics Authors: Bingxin Xu, Yuzhang Shang, Emilio Ferrara Published: 2026-08-17 arXiv: 2608.16889

Introduction

Long-horizon robot manipulation chains many contact-rich skills into one multi-stage task. Vision-language-action (VLA) models increasingly master individual skills, yet the chain still fails: errors compound beyond the policy's ability to correct, and one subtask silently constrains the next.

The Problem with Agentic Orchestration

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. Exponential exploration cost: Competence comes from whole-task exploration at test time. 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. Missing transition representation: VLA primitives have exit conditions but no 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 them. Cost becomes additive (T*K), and every failure is attributable to a single stage.
Transition-aware memory:
*Verification transitions*: the agent controls when to invoke the transition.
*Handover transitions*: restore the entry state perturbed by the predecessor.
*Lookahead transitions*: select policies whose outcomes the successor can inherit.

No parameters are updated anywhere in the pipeline.

Results

On the long-horizon benchmark RoboMemArena, BATON improves over state-of-the-art baselines by:

+11.6% task success rate
+14.9% cumulative success rate