A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a novel method for training robots using data from multiple smaller, less capable robots. The research addresses a key challenge in robotics: expensive, high-performance robots are often required to generate the large datasets needed for effective machine learning. The team’s ‘policy composition’ technique allows a larger ‘student’ robot to learn from the collective experiences of multiple smaller, cheaper ‘teacher’ robots, each with different physical capabilities and limitations. This approach successfully enabled a large two-fingered robot to learn complex manipulation tasks, like rotating a mug, from data gathered by several smaller single-fingered and two-fingered robots. The method could significantly reduce the cost and complexity of gathering training data for advanced robotic systems, making sophisticated robotic learning more accessible. For the full details, read the complete article at https://technologyreview.com/2024/07/18/1094825/robot-training-multi-robot-data/.