BATON: Long-Horizon Robotic Manipulation via Intelligent Subtask Exploration and Transition-Aware Memory

Background and Context

In the field of robotic manipulation, long-horizon tasks involve a series of complex, contact-intensive skills that must be executed in precise sequence to achieve a final goal. While current Vision-Language-Action (VLA) models demonstrate high proficiency in executing single, isolated skills, they frequently fail when tasked with multi-stage, long-chain operations. This failure is primarily driven by two systemic issues: error accumulation, where minute deviations in early steps amplify throughout the task chain, and subtask disconnection, where the terminal state of one action fails to satisfy the required initial conditions of the subsequent one.

Existing methodologies often attempt to resolve these issues by utilizing Large Language Models (LLMs) as high-level planners. In these setups, the LLM handles linguistic planning while VLA models are triggered only during contact-heavy phases. However, this approach faces a significant hurdle regarding exploration costs, which grow exponentially as the number of task stages increases. Furthermore, because these models lack an explicit representation of task transitions, the continuity of the task chain often collapses at the junction between different subtasks. The BATON framework was developed specifically to address these bottlenecks by restructuring how robots explore and remember task transitions.

Deep Analysis

The core technical innovation of the BATON framework lies in its shift from full-task exploration to intelligent subtask exploration and transition-aware memory. To combat the exponential growth of exploration complexity, BATON treats each subtask as an independent exploration unit. Instead of attempting to learn an entire long-horizon sequence at once, the system explores individual subtasks within short-horizon environments. Successful solutions are then stored in a memory bank. This architectural change reduces the complexity of exploration from an exponential scale to a linear scale, allowing the system to pinpoint exactly which stage of a task sequence has failed.

To manage the continuity of these tasks, BATON introduces a transition-aware memory system comprising three specialized mechanisms: Verifier Agents, Handoff Transitions, and Lookahead Transitions. The Verifier Agent utilizes wrist-view camera data to confirm that the environment has reached the necessary state to trigger VLA execution. The Handoff Transition mechanism is designed to correct residual environmental effects left by a previous task, ensuring the state is reset to the required starting point for the next step. Finally, the Lookahead Transition mechanism allows the agent to predict the outcomes of various strategies, selecting the execution path that is most compatible with subsequent tasks. Notably, these improvements are achieved without updating any underlying model parameters, relying instead on intelligent retrieval and logical reasoning.

Industry Impact

Experimental validation of the BATON framework was conducted using the RoboMemArena benchmark, which tests complex, multi-step manipulation sequences. The results demonstrate a significant leap in performance over current state-of-the-art (SOTA) methods. Specifically, BATON achieved an 11.6% increase in the task success rate and a 14.9% improvement in the cumulative success rate. These metrics underscore the framework's ability to maintain robustness across extended operational periods where traditional VLA models typically succumb to cumulative error.

By decomposing long-horizon tasks into modular, independently searchable units, BATON provides a viable technical path to solving the "combinatorial explosion" problem in robotic learning. The ability to achieve these gains through memory and logic rather than massive parameter scaling offers a more efficient trajectory for developing embodied intelligence. For the robotics industry, this suggests a move toward systems that can handle multi-step, high-precision production processes with greater flexibility than current single-skill specialized robots.

Outlook

BATON offers a new paradigm for the field of Embodied AI by proving that task organization and memory retrieval optimization can significantly enhance performance without increasing the underlying model's parameter count. This approach shifts the research focus from merely increasing the scale of VLA models to improving the structural logic of how agents interact with their environments over time.

As VLA models continue to advance in their fundamental capabilities, the development of architectural frameworks like BATON—which provide the "brain" necessary for high-level planning and precise task衔接 (connection)—will become a critical direction for achieving general-purpose robotic intelligence. This evolution marks a transition from simple skill learning toward true task understanding and execution, paving the way for autonomous agents capable of complex, long-term reasoning in real-world settings.