Vision-language-action (VLA) models are promising for field robotics, but their deployment is limited by strict edge constraints on size, weight, power, and cost. Compression techniques such as quantization, distillation, and pruning are often used to fit these models onto resource-limited hardware. A new arXiv paper focuses specifically on pruning and its effects on VLA models.
The paper's title, "Don't Run with Scissors," signals its central finding: pruning breaks VLA models. The abstract states that pruning degrades these models, but it also notes that the damage is not permanent—the models can be recovered. The authors do not appear to compare pruning with the other techniques in this snippet, so the claim is specific to pruning.
The result matters because pruning is a standard compression method, and if it silently harms VLA performance, robotics engineers need to know. The recovery aspect is the encouraging part: it suggests that the negative effects are not inherent to pruning itself, but can be undone with appropriate post-hoc steps. The paper does not specify the recovery method in the abstract, but the finding opens a practical path for deploying VLA models under edge constraints without sacrificing reliability.