Reproduce “ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models” on OpenResearch

About this paper

Ye Li, Huanan Liu, Kangye Ji, Yuan Meng, Jiajun Fan, Yuansong Wang, Shiyu Qin, Chenglei Wu and 2 more

Researchers from Tsinghua University and the University of Illinois at Urbana-Champaign developed ElegantVLA, a framework that employs phase-adaptive dynamic compute scheduling to significantly enhance the inference efficiency of Vision-Language-Action models. It achieves up to 3.77x speedup in simulation and 2.18x in real-world tasks, improving control frequency and often increasing task success rates.

Key insights

• Computational demands for VLA models are highly non-uniform across task phases, making dynamic, phase-adaptive scheduling a more effective strategy than static acceleration.
• Decoupling semantic stability (probed via CKA) from low-level motion continuity (robot speeds) provides complementary signals for coordinated and efficient compute reuse decisions across the Vision-LLM backbone and action head.
• A two-stage RL training approach, progressing from rule-guided initialization to task-efficiency optimization, is critical for learning a robust scheduler that effectively balances performance and computational cost.

Summary by alphaXiv.