SmolVLA: A Vision-Language-Action Model for…

SmolVLA compresses a vision-language-action policy to 450 million parameters and predicts fifty-action chunks with a flow-matching expert. Read it to see how an asynchronous queue overlaps inference with robot motion, why that makes execution faster, and where the reported success rate falls.

Reading focus: How a shortened vision-language backbone and interleaved action expert turn images, language, and robot state into action chunks. How a refill threshold lets a robot execute one chunk while a policy server predicts the next. How to read the compact-model results alongside narrow embodiment coverage and the measured speed-success tradeoff.

arXiv 2025. Mustafa Shukor et al.. 36 min read, medium difficulty.