From Executing Instructions to Autonomous Task Completion: RobotAgentBench Evaluates Open-World Robot Autonomy

RobotAgentBench is a new benchmark that evaluates the open-world autonomy of robots by simulating real-world family environments. It tests various aspects of robot intelligence, including task discovery, context reasoning, and robustness. The study compares different agent architectures and shows that combining code control with learning strategies significantly improves performance. The benchmark also highlights the challenges of transitioning from executing explicit instructions to autonomously completing user needs in dynamic environments.