Robots are getting smarter at performing complex, multi-step movements by combining artificial intelligence with principles from control theory. Researchers at Yale University, led by Assistant Professor Ian Abraham of mechanical engineering, have developed a new approach that enables robots to learn and execute advanced motor skills—such as flipping over and balancing into a handstand—by seamlessly integrating different learning methods. While AI techniques like reinforcement learning are effective for teaching robots single, specific tasks—like performing a backflip—challenges arise when trying to chain these skills together. “We often want to train our robots to learn new skills by compounding existing ones,” Abraham explained. “But AI models trained for multiple tasks tend to underperform compared to those trained on individual tasks alone.” To address this, Abraham’s team turned to hybrid control theory, a mathematical framework that determines the optimal moments for a robot to switch between different control modes during a task. This allows robots to transition smoothly between learning strategies—such as learning from trial and error (reinforcement learning) or using predictive models to plan movements—based on what’s most efficient and safe at each stage. The researchers tested their approach on a dog-like robot, successfully training it to perform a controlled flip and then stabilize into a balanced stance. AI was used to develop the complex, high-precision movements required for the flip, while hybrid control theory orchestrated the timing and coordination between different learning mechanisms. This integration ensures that the robot maintains performance quality even as it performs increasingly complex sequences. Abraham likens the process to how humans learn: “Think of how we learn new skills or play a sport. We first try to understand and predict how our body moves, then eventually movement becomes muscle memory and so we need to think less.” The goal is to help robots reach a similar level of intuitive, efficient performance. The work, published on the arXiv preprint server, could pave the way for robots to operate more effectively in unstructured environments like homes or disaster zones. “If a robot needs to learn a new skill on the job, it can draw from a range of learning methods—planning, reasoning, and experience—ensuring safety and success,” Abraham said. “Once it gains confidence, it can then use specialized, learned skills to go beyond basic tasks and perform at a higher level.”