The standard recipe for long-horizon robot manipulation—freeze the VLA, let an LLM agent plan in language, use analytic primitives for free-space motion, invoke the VLA only for contact-rich segments—breaks at scale in two ways. A new paper, BATON (arXiv 2608.16889), published August 17, 2026 from USC researchers, names both failures precisely and fixes them without updating a single model weight.
The first failure is combinatorial. When competence comes from whole-task test-time exploration and each stage needs T episodes to converge, a K-stage task requires roughly T^K total episodes. A failure anywhere in the chain yields no signal about which stage caused it. BATON reframes the problem: make the subtask the unit of exploration. Each subtask runs in a cheap short-horizon regime and stores its solution in memory. The long-horizon trajectory is assembled from stored solutions rather than discovered end-to-end. Cost collapses from T^K to T×K—additive rather than multiplicative in the number of stages—and every failure maps to a single subtask.
The second failure is structural. A VLA primitive carries an exit condition (task completion) but no entry condition (what the scene must look like before invocation is safe). A predecessor subtask can succeed in a form its successor cannot use—object grasped at the wrong angle, a container left slightly open. BATON calls this transition contamination. It addresses the problem with a transition-aware memory layer spanning three mechanisms. Within a subtask, a verifier agent holds the VLA back until a wrist-view image confirms the scene is ready. Across subtasks, a handoff transition actively restores the entry state the predecessor disturbed. A lookahead transition selects among candidate strategies by simulating which outcome the next subtask can inherit.
On RoboMemArena—a 26-task benchmark where average trajectories exceed 1,000 steps and 68.9% of subtasks depend on memory—BATON improves task success by 11.6% and cumulative success by 14.9% over the prior state of the art. No parameters are updated. The gains come purely from restructuring when and how the VLA is invoked, not from model capacity.
| Metric | Value |
|---|---|
| Tasks in RoboMemArena | 26 |
| Average trajectory length | >1,000 steps |
| Subtasks dependent on memory | 68.9% |
| Task success improvement vs. prior SOTA | +11.6% |
| Cumulative success improvement vs. prior SOTA | +14.9% |
| Model parameters updated | 0 |
The architecture is deliberately modular. The LLM planner, VLA primitive, and verifier agent are each a distinct model invocation. That separation makes cost accounting tractable: expensive VLA calls happen only after the verifier clears the scene, so compute scales with verified-ready states rather than total timesteps. Language memory storing subtask solutions persists across a session, so a solved stage doesn't re-explore when the same task configuration appears again.
BATON does not address one limitation: the verifier itself can fail. The wrist-view check is a VLM call, and if it misreads a scene, the handoff transition may activate on a contaminated state. The paper evaluates on RoboMemArena's simulation tasks; the sim-to-real gap on the wrist-view verifier remains an open question. The lookahead mechanism also assumes subtask strategies are enumerable enough to compare, which may not hold for highly open-ended manipulation goals.
The pattern is not specific to robotics. Any agent pipeline chaining specialized model calls—a retrieval model, code interpreter, vision classifier, generative step—faces the same two failure modes: whole-chain exploration cost scaling exponentially with stages, and exit conditions leaving state the next model cannot consume. BATON's answer: decompose exploration to the subtask level, gate every expensive model call behind a cheap verifier, and make entry/exit conditions explicit in memory. Cost goes linear; failures localize.