Don't Drop the BATON: Long-Horizon Robot Manipulation via Agentic Subtask Exploration and Transition-aware Memory
# BATON: Handling Long Robot Tasks Through Better Planning
Researchers have developed a system called BATON that addresses a fundamental problem in robot manipulation: while modern AI models can perform individual tasks well—grasping, placing, rotating objects—they struggle when these actions must be chained together over many steps. The approach uses a frozen vision-language-action (VLA) model for executing individual movements, while placing a large language model (LLM) agent in a supervisory role to plan the sequence of subtasks. The LLM reasons in language about what should happen next, then translates those decisions into robot actions, introducing a planning layer between intention and execution.
The core insight addresses a practical constraint in multi-stage operations: errors in early steps (a slight misalignment during pickup) don't just fail immediately—they propagate invisibly into later steps, where the robot cannot adapt or recover. By having an LLM actively manage transitions between subtasks and maintain awareness of task state, the system can catch or work around these compounding errors. For logistics, manufacturing, and field operations, this matters because real work rarely consists of single, isolated actions; most valuable tasks require 5-10+ sequential steps with dependencies.
Operators should note that this architecture separates planning from low-level control, which has practical consequences: it requires clear state observation between steps and benefits from explicit task specification, but it may reduce real-time reactivity compared to end-to-end learned models. The approach trades some autonomy (the LLM follows a plan) for interpretability and correction capacity—a tradeoff worth monitoring as these systems move into production environments.