You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
[EMNLP 2026] MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research · 浏览器里运行的安卓模拟器 · Browser-hosted Android Simulator · Verifiable Evaluation · Scalable Online RL Training
Official implementation of COST-HIL: Cost-Aware Human-in-the-Loop Reinforcement Learning for Real-World Dexterous Manipulation. This repo contains source codes, experiment logs, real-robot test scripts and supplementary materials for our CVPR submission.
A tutorial and resource hub for Flow Matching in Reinforcement Learning (FM-RL), accompanying our TMLR survey. Includes an introduction to FM-RL, a curated collection of papers, method taxonomy, code repositories, and resources for researchers and newcomers to the field.
A progressive hands-on lab series comprising 16 post-training experiments, covering SFT, parameter efficient fine-tuning with LoRA & QLoRA, adding task-specific head for binary sentiment classification and NER, and three RL methods (PPO, GRPO, and DPO) to align a SFT-tuned model with human preference for a conditional sentiment continuation task.
Online goal-reaching RL with diffusion planning: a 2D point-mass prototype plus a Maze2D Diffuser workflow driven by an autonomous agentic experiment controller.