A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in robotic manipulation. Researchers have developed a system that allows a robot to learn complex, dexterous manipulation tasks, such as spinning a long baton, using a relatively small amount of real-world data. The key innovation is a two-stage learning framework: the robot first learns a general understanding of physics and object interaction from a massive, offline dataset of simulated examples. This pre-trained model is then fine-tuned with a few hours of physical practice, enabling it to adapt to the real world’s complexities and unpredictability. This approach marks a shift from methods requiring vast amounts of costly real-world trial-and-error, potentially accelerating the development of robots capable of nuanced, human-like tasks in unstructured environments. For the full details, read the article at https://technologyreview.com/2024/05/20/1093035/this-robot-can-spin-a-better-pen-trick-than-you/.