ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning
技术动态
来源：arXiv cs.RO发布时间待核实
arXiv:2608.19182v1 Announce Type: new Abstract: We introduce Accelerating Dexterity via Pre-Training (ADEPT), a large-scale reinforcement learning (RL) framework for learning sim-to-real transferable dexterity across high degree-of-freedom (DoF) robot embodiments that can solve long-horizon tasks directly from raw visuo-tactile perception. ADEPT pretrains a dexterous policy on a generic object reposing task, then post-trains downstream policies with this pretrained behavior as a prior.