EVENT DOSSIER
On September 7, 2026, arXiv cs.LG published the Squint visual reinforcement learning method. This method can converge on the ManiSkill3 SO-101 task set (including eight operational tasks and strong domain randomization) in just 15 minutes on a single RTX 3090 GPU, with most tasks converging in less than 6 minutes. Squint significantly improves the training speed of wall clocks through mechanisms such as parallel simulation, distributed reviewers, resolution scaling, layer normalization, optimized updates and data ratios, and optimization, and successfully verified its ability to migrate from simulations to real SO-101 robots.
Squint 是一种视觉强化学习方法，在单张 RTX 3090 GPU 上仅需 15 分钟即可训练出在 ManiSkill3 SO-101 任务集（含八项操作任务及强域随机化）中收敛的策略，多数任务收敛时间不足 6 分钟。该方法通过并行模拟、分布式评论家、分辨率缩放、层归一化、调优的更新与数据比率及优化实现等机制，实现了比先前视觉离线和在线方法更快的墙钟训练速度，并成功验证了从仿真到真实 SO-101 机器人的迁移能力。