Pranit Mohnot Rose Hills
Impact Uncertainty Robust Hybrid DDP for Legged Robots
Hybrid Differential Dynamic Programming (DDP) is a control method for legged robots that is optimized for efficient motion. However, it is sensitive to modeling errors (the difference between how we expect the robot to behave and how it actually behaves) and unexpected disturbances (such as a sudden wind gust or a slippery floor). There exist DDP algorithms that are robust to some disturbances, but not those that occur when the robot’s foot makes contact with the ground (known as impact). Uncertain impact dynamics are currently a major source of disturbances in legged robot trajectory, which can lead to instability and falls. Many current solutions rely on machine learning to fine-tune controls, but machine learning does not provide the same performance guarantees as a more formal mathematical method.
To address these issues, I will derive equations (and approximations, for computational efficiency) that take into account possible disturbances. By finding the worst-case scenario at each impact, we can incorporate this information into the DDP framework and optimize the controller under the assumption that the worst-case happens. This will lead to a DDP controller that is robust to impact uncertainty and can maintain stability and prevent falls even in uncertain environments.