ThinkAct improves robot manipulation by combining high-level reasoning with low-level action execution. The sim-and-real policy co-training approach bridges the sim-to-real gap. RobotSmith leverages vision-language models to automatically design task-specific tools for robots.
