UCSD Hao AI Lab 发布 ORA v0.1,提升 RUA 后训练冷启动表现
ORA gives RUA post-training a cold start and a free win. On RoboLab-120, base Qwen3.8-27B scored 5.4% in PEAR’s setup and 52.5% in ORA. PEAR’s RL checkpoint rea
AI 导读
UCSD Hao AI Lab 发布零样本机器人使用开放运行框架 ORA v0.1,为 RUA 后训练提供冷启动收益。在 RoboLab-120 上,基础 Qwen3.8-27B 在 PEAR 设置中得分为 5.4%,在 ORA 中达到 52.5%;PEAR 的 RL checkpoint 在 ORA 中达到 61.9%。
正文
ORA gives RUA post-training a cold start and a free win. On RoboLab-120, base Qwen3.8-27B scored 5.4% in PEAR’s setup and 52.5% in ORA. PEAR’s RL checkpoint reached 61.9% in ORA. https://t.co/4reWMSYDsm
AI 热点