TS-DP: Reinforcement Speculative Decoding for Temporal Adaptive Diffusion Policy Acceleration
Published in IEEE Transactions on Multimedia (TMM 2026)
Recommended citation: Ye Li, Jiahe Feng, Yuan Meng, Kangye Ji, Chen Tang, Jiajun Fan, Xinwan Wen, Jiawei Li, Wenwu Zhu, Shu-Tao Xia, Xin Wang, Zhi Wang. "TS-DP: Reinforcement Speculative Decoding for Temporal Adaptive Diffusion Policy Acceleration." IEEE TMM 2026. https://arxiv.org/abs/2512.15773
Speculative decoding for diffusion policies. Diffusion policies in embodied control are I/O-bound rather than compute-bound, so TS-DP distills a Transformer drafter that imitates the base policy’s costly denoising calls, and an RL scheduler adjusts the speculative parameters as task difficulty changes over time. In simulation it cuts function evaluations by up to 4.17× and wall-clock time by up to 3.2× with a draft acceptance rate above 85%; on real robots it cuts evaluations 3.8× and runs at 26 Hz, raising success on real-time tasks such as a conveyor belt.
