DescriptionUnderactuated robots are mechanical systems with fewer control inputs than their degrees of freedom (DOF). Inverted pendulums, bicycles and walking robots are a few examples of such systems. Underactuated balance robots are underactuated robots that must perform the balancing and tracking tasks simultaneously. The balancing task requires the robot to maintain its balance around possibly unstable equilibrium points, while the tracking task requires it to track desired trajectories. For these competing tasks, a common guideline to design controllers is to identify a low dimensional subspace of the state space, called a latent manifold, on which the balancing and tracking tasks are consistent and compatible. The approach of latent manifold identification is called model reduction. Previous works apply model reduction to physical-principled models with well understood dynamics structures. This dissertation proposes machine learning-based model reduction approaches for modeling high dimensional robots, extracting balancing skills from demonstration data and
controlling robots with data-driven models.
Several aspects of machine learning make it attractive for use in model reduction and control applications for underactuated robots. First, the system dynamics learned from collected data can be more accurate than analytical models derived from physical laws. For high dimensional motion, it is also easier to build learning-based latent space models than any physical models. Second, a latent manifold that encodes balancing skill can be learned from the demonstrated trajectories. One application example is to transfer human walking skills to humanoid robots by enforcing humanoid robots onto the latent manifold identified from human trajectories. Finally, optimization-based controllers such as model predictive control (MPC) and reinforcement learning can be integrated with the learned model and applied to stabilize the learned open-loop dynamics onto the desired latent manifolds.
In this dissertation, we introduce a framework that integrates the physical-based robot model with the learning-based latent manifold model for high dimensional human limbs motion to achieve pose estimation of human-robot interactions. Human-bikebot interaction is used as an example to demonstrate the proposed approach. We extend the physical latent manifold-based controller to achieve biped slip recovery. We reveal the relationship between learning-based model reduction and physical-based model reduction for high dimensional dynamics such as human legged locomotion. One of our final goals is to design learning-based controllers to achieve biped walking and slip recovery. The balancing while tracking problem has been successfully solved by designing physical model-based controllers to stabilize the system state onto the balance equilibrium manifold (BEM). However, its application has been restricted to systems with well understood dynamics structures. In the last part of this dissertation, we adopt the BEM concept to design a learning model-based control framework. The system dynamics is identified without prior physical knowledge nor successful balancing demonstrations. The proposed framework achieves superior control performance compared to the physical model-based approach, and provides analytical performance guarantees. The works in this dissertation are demonstrated using multiple robotic platforms such as an inverted pendulum, a bikebot and a biped robot.