BICPO-VLA framework cuts robot handoff delays with behavior-aware optimization

The request-to-handoff gap in async VLA control has three sources: ambiguity about the intended behavior at request time, physical-state drift while actions are generated, and residual incompatibility when the new action takes over. BICPO-VLA addresses each in sequence. First, an instruction-aware causal history encoder identifies the behavior supported by the command and current task progress, which anchors the rest of the pipeline. Second, sequential Haar subspace generation decomposes each action chunk into complementary scaffold and residual coefficients, letting the model generate those components separately and reconstruct them exactly. By reducing iterative refinement in the original action space, the approach shortens the interval during which the robot continues moving before a new chunk becomes available—directly shrinking the handoff lag.

Finally, the BICPO roll-out mechanism rolls known outgoing actions to the actual handoff state and applies reference-relative Flow-DPO among behaviorally matched candidates. This adapts the generated chunk to whatever mismatch remains without changing the intended behavior of the action. The result is asynchronously generated actions that hand off feel like a seamless continuation, rather than a jarring switch. The paper, submitted to arXiv (2608.13924), includes nine pages and four figures and focuses on robustness in robotic manipulation. While no benchmark numbers are given in the abstract, the architectural contributions target a real bottleneck in real-time VLA deployment: the gap between command and control.