Research paperRobotics & Embodied AI · Efficiency & Inference · Evaluation & Benchmarks2 sources · Oct 6, 2026

CARE: Certifying Acceleration for Vision-Language-Action Inference

Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically evaluating based on latency and average task success.

Qwen Vision-language-action models meta-llama/Llama-3.1-8B Qwen/Qwen3.5-9B

← Back to brief

Key points

While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive.
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.
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.

Sources (2)

[1]CARE: Certifying Acceleration for Vision-Language-Action Inference
Hugging Face Daily Papers · Oct 6, 12:00 AM
Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically evaluating based on latency and average task success.
While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive.
[2]CARE: Certifying Acceleration for Vision-Language-Action Inference
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 6, 06:00 PM · same content

Extractive summary: sentences quoted from the sources.