Event

FastVLA Visual Token Compression Framework Accepted to WACV 2027

1 reports · 1 sources

The FastVLA author team proposes a training-free, task-aware, and temporally adaptive visual token compression framework to improve the closed-loop inference efficiency of vision-language-action (VLA) models; the work has been accepted to WACV 2027. The framework scores tokens via task-to-vision attention and adaptively sets the compression rate based on the uniformity of the importance distribution (calibrated using the inverse Simpson index and an online sliding window). The authors evaluated it on OpenVLA, OpenVLA-OFT, and CogACT across the LIBERO and SIMPLER simulation environments and on a physical Franka Research 3 robot: in simulation, LLM backbone FLOPs were reduced by 35-59%, action latency was reduced by 9-32% (1.09×-1.46× speedup), and task success rates remained unchanged; on the physical robot, action prediction latency was reduced by 20% (1.25× speedup), with comparable success rates.