Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Yu Chao reports CEL training with up to 96 robots

AI summary

In a DeepTech interview, Tsinghua researcher Yu Chao described her team’s Collective Experience Learning (CEL) method, which lets robots learn from successful experiences collected by other robots. In a phone-packing task, success rose from 86.7% with 16 training robots to 98.3% with 96. The final test used robots that had not participated in training and consisted of 120 trials, with two failures, according to Yu. The team also described RLinf, a large-scale reinforcement-learning framework, and Harness VLA, which combines a Coding Agent for high-level planning with different robot-control strategies and external memory. Yu said future robot learning may increasingly rely on cloud-based models and retrieved experience rather than only changing model parameters through training.

In-depth insight

The results provide quantitative real-robot evidence that CEL performance increased with the number of training robots in the tested phone-packing task. The researchers also used a held-out-robot test and a controlled evaluation process, while emphasizing that the findings may not generalize beyond the tested task, robot scale and homogeneous robots.

2022 年,一篇看似没有提出复杂新算法的论文,成为多智能体强化学习领域最有影响力的工作之一。

这篇论文叫 MAPPO。研究者发现,只要把已有的强化学习算法 PPO 用好,调整训练和实现中的一些细节,就能在多智能体协作任务中取得相当出色的效果。这个结论如今看起来并不难理解,但在当时,学术界更看重新算法和理论创新,MAPPO 因为方法过于简单,经历了多次投稿受挫。

论文的主要作者于超,当时还是清华大学电子工程系的一名博士生。如今,这项工作的引用量已接近 5,000 次,成为该领域广泛使用的基准方法之一。于超目前是清华大学深圳国际研究生院助理教授、博士生导师,同时也是物理智能公司正行创新(Striding AI)的联合发起人。

#Collective Experience Learning (CEL)#MAPPO#PPO#RLinf#Harness VLA#Coding Agent#VLA#Tsinghua University#Striding AI#Yu Chao

Original · 24 min readRead the original
Group chat

No comments yet. Start the conversation.