0.1% Tsinghua AI Researcher Gives Paper Ideas For Robot RL

Next Sunday Skool LIVE: Build LLM In 1 Prompt - Get live 1-on-1 help in a weekly group call - write papers, train LLMs: https://www.skool.com/become-ai-resea... arxiv - https://arxiv.org/abs/2604.08508 📆 Schedule 1 on 1 with me - https://cal.com/vuk-ai/60-min A Tsinghua robotics researcher (PhD in Hong Kong) walks through a paper that combines sampling-based MPC with a reinforcement learning controller for robot manipulation. The idea: keep RL for low-level control, then stack a sampling-based planner on top to actually complete high-level tasks like moving a desk or making coffee. He also argues the architecture shouldn't be static — you can distill the sampling planner back into a policy, so RL and sampling blend into one "fluid" system. 0:00 A Tsinghua researcher on robotics RL 0:31 RL is great at low-level control, not whole tasks 1:04 Put a sampling-based MPC planner on top 1:40 RL vs classical planning: which part goes where 2:27 Distill sampling into a policy (a fluid architecture)