It is amazing what you can learn if you're not afraid to try.

PPO GRPO DAPO GSPO SAPO GDPO

SimpleVLA-RL 1. 研究背景 Simple VLA-RL: scaling VLA Training via Reinforcement Learning: https://github.com/PRIME-RL/SimpleVLA-RL 问题：针对机器人领域VLA模型主要采用大规模预训练和有监督微调（supervised fine-tune）的方案，存在两个问题...

guide line 热力学扩散模型 DDPM DDIM Flow Matching B站：扩散模型基础 1. 扩散模型基础 扩散模型：模拟热力学扩散过程来构建的图片生成式神经网络模型。 高维空间中图片的表示 高纬向量空间中图片的表示：每一张图片对应的就是高维向量空间中的一个点。 郎之万动力学方程Langevin Equation 郎之万动力学方程（Langev...

Domain randomization for transferring deep neural networks from simulation to the real world https://readpaper.com/pdf-annotate/note?pdfId=4557557081012445185&noteId=3173811748193745664

参考文档 https://www.doubao.com/thread/w1b0ce4572028377f BCQ IQL TD3+BC 1. 离线强化学习 2. CQL: Conservative Q-Learning for Offline Reinforcement Learning https://readpaper.com/pdf-annotat...

参考文档 #GEN-0 / Embodied Foundation Models That Scale with Physical Interaction

针对策略梯度优势估计的问题「高方差variance、高偏差bias」提出统一框架GAE来平滑地处理variance和bias之间的平衡。

参考文档 # Transformer-based Multi-Agent Reinforcement Learning for Generalization of Heterogeneous Multi-Robot Cooperation

A framework that abstracts RL as a sequence modeling problem.