BATON architecture decomposes long-horizon robot tasks into phased exploration and transition-aware memory

Robots face a complex operational challenge when executing long-horizon interactive tasks that require chaining contact-rich skills across a multi-stage trajectory. While vision-language-action (VLA) models master individual skills with high efficiency, the operational chain breaks down when combining them, as errors accumulate beyond the motion policy's ability to correct and subtasks constrain subsequent stages without clear indicators.

The current prevailing approach relies onfreezing the weights of the vision-language-action model and delegating control to a large language model agent, where the agent handles natural language planning and navigates free space through analytic primitives, restricting VLA model invocation to segments involving intensive physical contact while logging adaptations in linguistic memory. However, this architecture runs into two fundamental flaws when generalized to long horizons: the first is a reliance on full-task exploration at test time, an approach whose exploratory cost scales exponentially with the number of stages, such that if a single stage requires T attempts, a K-stage task requires roughly T^K attempts, alongside an inability to pinpoint the specific stage causing a failure. The second flaw is the complete absence of transition representation, as VLA primitives carry exit conditions without prior entry conditions, leading to subtasks succeeding in terminal states that subsequent stages cannot handle.

To address these two dilemmas, researchers Pengxin Xu, Yuzhang Shang, and Emilio Ferrara introduced a new architecture called BATON. In addressing the first flaw, the framework establishes the subtask itself as the primary unit of exploration, allowing each subtask to be explored independently within a short, low-cost horizon before storing its standalone solution in memory. Instead of conducting an exploratory process across the entire long trajectory, the full task trajectory is composed from these ready-made solutions, reducing the exploration cost from an exponential relationship to a linear additive formula of T*K, and allowing any occurring failure to be attributed solely to its relevant stage.

To counter the second dilemma, the researchers equipped the exploration system with a transition-aware linguistic memory that operates through integrated pathways: within a single subtask, a verifier agent manages the invocation transition, invoking the VLA model only after wrist-camera vision confirms the readiness of the scene and position. Across subtask boundaries, a handover mechanism intervenes to reset entry conditions that might be disrupted by residual effects from the preceding stage, while a lookahead mechanism selects strategies that yield outputs the subsequent stage can inherit and build upon successfully.

This architecture operates entirely without modifying or updating the weights of the foundation models. Experiments on the long-horizon benchmark RoboMemArena showed marked superiority for the BATON system, achieving an 11.6 percent increase in task success rate and a 14.9 percent improvement in cumulative success rate compared to the strongest prior baselines.