Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think

2026-06-18

AI Summary

This work addresses the high computational cost of fine-tuning and inference in large-scale vision-language-action (VLA) models. It reveals, for the first time, significant inter-layer representational redundancy in pretrained VLA models and introduces a training-free depth compression method. By leveraging centered kernel alignment (CKA) to analyze layer similarity, the approach identifies and removes redundant layers through a single forward pass. The resulting model achieves 50% depth reduction while matching or exceeding the original model’s performance across ten real-world robotic tasks and simulated environments. Moreover, it reduces fine-tuning time by 40–50% and accelerates real-time inference by 30%, establishing a new paradigm for training-free compression of VLA models.

Abstract

Vision-Language-Action (VLA) models pre-trained on massive video-robot datasets have revolutionized robotic manipulation, yet their multi-billion parameter architectures impose prohibitive computational burdens during downstream fine-tuning and real-time inference. In this work, we reveal a highly non-trivial architectural characteristic of these continuous control foundation policies (e.g., pi_0, GR00T-N1.5): despite being trained on diverse physical trajectories, they exhibit severe layer-wise representational redundancy. To exploit this, we introduce a structural compression pipeline that is entirely training-free, bypassing the need of existing methods to load full-scale models to learn optimized token reductions or dynamic layer selectors. Instead, using only a single forward pass via Centered Kernel Alignment to identify redundant layer features, we remove twin layers to permanently compress the model depth by up to 50% across both the VLM backbone and the continuous control policy head. Downstream fine-tuning of this streamlined architecture yields a dual acceleration benefit: a 40-50% reduction in training time and up to 30% faster real-time inference, while matching or exceeding full-scale base model performance. We comprehensively validate our method across three simulation benchmarks (LIBERO, RoboCasa, SimplerEnv) and 10 diverse real-world manipulation tasks across 4 unique robotic embodiments. These results prove that advanced VLAs require significantly fewer layers than previously assumed, offering a highly compute-efficient paradigm for scalable robot learning.

Problem

• Vision-Language-Action models
• computational burden
• fine-tuning
• real-time inference
• model compression

Innovation

• structural compression
• layer redundancy
• training-free compression
• Vision-Language-Action models
• Centered Kernel Alignment