Don't Drop the BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory
BATON is a method for long-horizon robot manipulation that uses an LLM agent with a frozen vision-language-action (VLA) model. It addresses two failure modes: multiplicative exploration cost in multi-stage tasks and lack of transition representation between subtasks. BATON makes the subtask the unit of exploration, storing solutions in memory, and composes long-horizon trajectories from these solutions.
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 because 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; (2) it has no representation of transitions, so a subtask can succeed in a form its successor cannot use. BATON addresses these by making the subtask the unit of exploration, storing solutions in memory, and composing long-horizon trajectories from these solutions.
BATON introduces subtask-level exploration to avoid the T^K cost of whole-task exploration, and transition-aware memory to capture entry conditions between subtasks. The approach freezes the VLA and uses an LLM agent for planning and memory, which is a notable architectural choice for long-horizon manipulation.
This research targets a key bottleneck in deploying robot manipulation in industrial and service settings: reliability over long multi-step tasks. By reducing exploration cost and improving transition handling, BATON could make VLA-based robots more practical for real-world automation.
If BATON proves effective, it could lower the cost and time to deploy long-horizon manipulation skills, increasing the commercial viability of robot automation in logistics, manufacturing, and household robotics.
Next signals to watch include whether BATON is evaluated on real robot hardware, whether the subtask memory generalizes across tasks, and whether the approach is adopted by robotics companies or integrated into existing VLA frameworks.