BAAI 联合多所高校推出 SWITCH 基准，测试真实环境中的闭环界面操控
BAAI 联合上交大等几个机构做了个 SWITCH 基准，专门考 AI 能不能按电梯、开开关这类闭环操作，和以往只测看图的传统基准不一样。
SWITCH（Semantic World Interface Tasks for Control & Handling）由北京通用人工智能研究院（BAAIBeijing）联合上海交通大学、中国科学院自动化所、武汉人工智能研究院等机构共同推出。基准面向开关、家电面板、电梯等实体控制接口（TCIs），测试具身智能体在第一人称真实环境中的交互推理。不同于以往只测开环感知或单步动作的具身基准，SWITCH 强调执行、反馈与纠错的持续闭环。
Switches, appliance panels, elevators, and embedded GUIs are the "Physical APIs" of our daily environments. Yet most embodied benchmarks only test open-loop perception or single-step actions—ignoring the continuous loop of execution, feedback, and error recovery.
Introducing SWITCH(Semantic World Interface Tasks for Control & Handling), a comprehensive benchmark for closed-loop interactive reasoning with Tangible Control Interfaces (TCIs) in real-world egocentric environments, jointly presented by @BAAIBeijing, SJTU, UCAS, CASIA, Wuhan AI Research, and HIT.
