RobotAgentBench comprehensive evaluation of robots' open-world capabilities

RobotAgentBench evaluates robots' open-world capabilities, requiring robots to understand natural language requests and dynamically adjust plans. The evaluation includes multi-scenario tasks, testing whether robots can recognize object changes and adapt to environmental adjustments. Results reveal that current robots struggle to handle complex natural language and dynamic environmental interactions.