Completion at the Boundary (CaB): Deployable Switching with Completion-Aware Control under Limited Calibration

Researchers propose Completion at the Boundary (CaB), a novel approach for vision-language-action agents to determine when to switch between sequential instruction steps without requiring test-time relearning. The method uses Boundary-Phase Tokens to preserve two-sided evidence for completion decisions, improving composite task execution in robotic control systems.

This research addresses a critical gap in deploying vision-language-action agents: determining when one instruction completes and another should begin. In multi-step tasks like "do A, then B," mistimed handoffs create cascading failures downstream. Traditional approaches collapse completion signals into single scalar values, which become brittle when task characteristics shift—a problem inherent to any deployment system operating under low-calibration constraints.

The core innovation lies in Completion at the Boundary's two-pronged design. Rather than forcing asymmetric boundary evidence into a single decision point, CaB predicts event-local completion objects using Boundary-Phase Tokens (Before/Hit/After), preserving nuanced context. CaB-When converts this richer representation into minimal, auditable switching decisions, while CaB-How reuses the same completion object to condition action generation for stable control during handoffs. This architecture mirrors how human operators handle task transitions—maintaining situational awareness while executing the next step.

The research validates this approach using an intervention-aware E1/E2 protocol on a first-person Minecraft VLA benchmark, demonstrating improvements in composite execution and handoff quality under matched capacity constraints. This matters because it sidesteps the expensive requirement for test-time relearning or multiple task-specific calibration procedures, making deployment more practical. The constraint of using a single globally calibrated switching rule reused unchanged across the test set reflects real-world operational requirements where dynamic recalibration isn't feasible.

For AI development teams building embodied agents, this research offers a deployable pathway to more reliable task composition without architectural redesigns. The focus on auditable decisions and boundary-stable control addresses practical concerns in autonomous systems where failures can be costly.

• →CaB enables reliable task switching in vision-language-action agents without test-time relearning or per-task calibration
• →Boundary-Phase Tokens preserve two-sided completion evidence, improving robustness across varying task characteristics
• →The method demonstrates improved composite task execution on Minecraft benchmarks with matched capacity constraints
• →Design emphasizes auditable switching decisions, critical for safe deployment of embodied AI systems
• →Approach addresses practical deployment requirements where global calibration must remain fixed across diverse test scenarios