CARE Adds a Risk Gate to Vision-Language-Action Acceleration

Introduction

Vision-language-action (VLA) models are becoming an important interface between visual understanding, language instructions, and robot control. Their main deployment challenge is cost: a policy may need to run at every control step, making inference expensive and limiting responsiveness. Action chunking, visual-token pruning, and fewer generation steps can reduce this burden, but they may also remove information that matters at a particular point in a trajectory.

Average task success and mean latency do not fully expose that trade-off. An accelerated policy can look competitive overall while occasionally breaking a task that the original policy would have completed. In closed-loop control, even a small action difference can compound over many steps, so the relevant failure may only become visible at the end of an entire episode.

What CARE changes

• A more targeted failure definition. CARE counts an acceleration-induced failure when the reference policy succeeds but the accelerated candidate fails. This separates damage caused by acceleration from failures that both policies would have experienced.
• Paired rollouts from the same starting point. The reference and candidate are evaluated under matched initial conditions, with terminal episode outcomes used as the primary evidence. The procedure therefore works across different acceleration mechanisms without requiring a shared internal representation.
• Certification before deployment. Candidate accelerators are tested on a calibration set. CARE applies finite-sample statistical guarantees to determine whether the observed risk remains below a user-specified budget. If no candidate qualifies, it falls back to the reference policy.
• An affordable evaluation strategy. Sequential testing can stop once a candidate is clearly certified or rejected. Reference rollouts are also triggered when accelerated runs fail, avoiding unnecessary exhaustive comparisons.

Results and implications

With OpenVLA-OFT on four LIBERO suites, CARE certified 9.0–10.8× speedups while guaranteeing, at 95% confidence, that at least 85.8% of episodes solved by the reference policy were preserved. Under tight budgets, selectors without formal guarantees exceeded the budget in as many as 75% of trials, whereas CARE stayed within the specified constraint. Its sequential version used 78.9% fewer rollouts than exhaustive evaluation. The authors also report applications to flow-step reduction for pi0.5 and to Qwen3.5-9B and Llama-3.1-8B agents in Crafter.

The broader contribution is methodological. VLA acceleration is reframed as a constrained selection problem: maximize speed, but only after controlling the probability of losing capabilities already demonstrated by the reference policy. That framing is useful for robotics deployments where reliability, auditability, and explicit failure budgets matter more than a single benchmark average.

CARE is not an unconditional proof of safety. Its guarantee depends on the calibration set, the distribution of initial states, and the quality of the terminal success criterion. In practice, it is best understood as a statistically grounded deployment gate that helps teams choose an accelerator without hiding rare closed-loop regressions behind aggregate metrics.

Source: Hugging Face Daily Papers