CARE: Certifying Acceleration for Vision-Language-Action Inference
CARE certifies VLA inference accelerators so speedups stay inside a user-specified failure budget.
CARE selects the fastest vision-language-action accelerator that stays under a user-specified failure budget, using paired rollouts for finite-sample guarantees. On four LIBERO suites with OpenVLA-OFT it certifies 9.0-10.8x speedups while preserving at least 85.8% of reference-solved episodes at 95% confidence. Selectors without guarantees exceed tight budgets in up to 75% of trials, while CARE's sequential form uses 78.9% fewer rollouts. The method also covers flow-step reduction for pi0.5 and Qwen3.5-9B and Llama-3.1-8B agents in Crafter.
- Defines acceleration-induced failure using paired closed-loop rollouts
- Certifies 9.0-10.8x speedups on LIBERO with OpenVLA-OFT
- Preserves at least 85.8% of reference-solved episodes at 95% confidence
- Sequential testing uses 78.9% fewer rollouts than exhaustive evaluation
- Also covers pi0.5, Qwen3.5-9B, and Llama-3.1-8B
Full article244 words · extracted from huggingface.co · click to collapse
While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive. Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically evaluating based on latency and average task success. However, acceleration may discard information and break tasks the original policy would solve, a risk hidden by average metrics. Measuring these failures is challenging because action deviations compound over closed-loop trajectories, meaning task failure is only observable across full episodes. We therefore define an acceleration-induced failure via paired rollouts from identical initial conditions, tracking when the reference succeeds but the accelerated policy fails. To manage this, we introduce CARE, an approach for certified accelerator selection. CARE uses paired rollouts on a calibration set to provide finite-sample guarantees that acceleration-induced failure risk stays below a user-specified budget. It deploys the fastest certified candidate, falling back to the reference if none qualify. By relying only on terminal outcomes and measured compute, CARE applies unchanged across diverse acceleration mechanisms, while sequential testing and failure-triggered reference rollouts keep certification affordable. On four LIBERO suites with OpenVLA-OFT, CARE certifies 9.0--10.8times speedups while guaranteeing (at 95% confidence) that at least 85.8% of reference-solved episodes are preserved. Under tight budgets, selectors without guarantees exceed the budget in up to 75% of trials, whereas CARE stays within budget and its sequential form uses 78.9% fewer rollouts than exhaustive evaluation. CARE further generalizes to flow-step reduction for π_{0.5}, and to Qwen3.5-9B and Llama-3.1-8B agents in Crafter.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2610.08917