arXiv Artificial Intelligence· Roberto Capobianco (Sony Europe Ltd), Harm van Seijen (Sony Corporate Technology Center America, Inc), Nolan D. Bard (Sony Corporate Technology Center America, Inc), Neil Burch (Work done while at Sony), Fatima Davelouis (Sony Corporate Technology Center America, Inc), Josh Davidson (Sony Corporate Technology Center America, Inc), Alisa Devlic (Sony Europe Ltd), Yunshu Du (Work done while at Sony), Ishan Durugkar (Sony Corporate Technology Center America, Inc), Siddhant Gangapurwala (Sony Corporate Technology Center America, Inc), Daniel Hernandez (Sony Corporate Technology Center America, Inc), G. Zacharias Holland (Sony Corporate Technology Center America, Inc), Sahil Jain (Sony Corporate Technology Center America, Inc), Kenta Kawamoto (Sony Group Corporation), Raksha Kumaraswamy (Sony Corporate Technology Center America, Inc), Patrick MacAlpine (Sony Corporate Technology Center America, Inc), Dustin R. Morrill (Work done while at Sony), Declan Oller (Sony Corporate Technology Center America, Inc), Francesco Riccio (Sony Europe Ltd), Akanksha Saran (Work done while at Sony), Craig Sherstan (Work done while at Sony), Kaushik Subramanian (Sony Europe Ltd), Thomas J. Walsh (Sony Corporate Technology Center America, Inc), Samuel Barrett (Sony Corporate Technology Center America, Inc), Kizza N. Frisbee (Sony Corporate Technology Center America, Inc), Mady Govil (Sony Corporate Technology Center America, Inc), Johannes G\"unther (Sony Corporate Technology Center America, Inc), Varun R. Kompella (Sony Corporate Technology Center America, Inc), James A. MacGlashan (Sony Corporate Technology Center America, Inc), Maxwell Svetlik (Sony Corporate Technology Center America, Inc), Michael D. Thomure (Sony Corporate Technology Center America, Inc), Jaden B. Travnik (Sony Corporate Technology Center America, Inc), Kevin Waugh (Sony Corporate Technology Center America, Inc), Elahe Aghapour (Work done while at Sony), Florian Fuchs (Work done while at Sony), Andreanne Lemay (Work done while at Sony), Shruti Mishra (Work done while at Sony), Takuma Seno (Work done while at Sony), Peter Stone (Sony Corporate Technology Center America, Inc), Michael Spranger (Work done while at Sony), Peter R. Wurman (Work done while at Sony)·· 13 小时前AI 评分44
Sony 提出可教练智能体框架，在《地平线：西之绝境》与《GT赛车》中实现风格可控
Coachable Agents for Interactive Gameplay
AI 导读
Sony 研究团队提出“可教练智能体”框架，将通用价值函数近似器（UVFA）与精选训练场景、学习算法和数据增强结合，让用户在训练完成后仍能在运行时控制任务完成风格。该框架已在 AAA 游戏《地平线：西之绝境》《Gran Turismo》及开源人形行走测试域中验证，各智能体在满足主任务的同时对风格请求保持较强一致性。论文已被 TMLR 接收，源代码与概览视频已公开。
推荐理由模型分析
索尼团队把通用价值函数近似器（UVFA）与训练场景、算法和数据增强结合，提出可在训练后对智能体行为施加“风格”控制的框架，并在《地平线：西之绝境》《GT赛车》和开源人形机器人域验证。对做强化学习游戏AI和机器人控制的开发者，可参考其风格约束与主任务兼顾的基线设置；但证据限于论文演示，未给出跨域泛化或部署成本数据。
来源：arXiv Artificial Intelligence · arxiv.org