Iterative Learning Control for A1 and Aliengo Robot
We have been working on developing bounding and pronking gait controllers for two quadrupedal robots A1 and Aliengo from Unitree Robotics. Precise trajectory tracking for legged robots can be challenging due to their high degrees of freedom, unmodeled nonlinear dynamics, or random disturbances from the environment. A commonly adopted solution to overcome these challenges is to use optimization-based algorithms and approximate the system with a simplified, reduced-order model. Additionally, deep neural networks are becoming a more promising option for achieving agile and robust legged locomotion. These approaches, however, either require large amounts of onboard calculations or the collection of millions of data points from a single robot. To address these problems and improve tracking performance, this paper proposes a method based on iterative learning control. This method lets a robot learn from its own mistakes by exploiting the repetitive nature of legged locomotion within only a few trials. Then, a torque library is created as a lookup table so that the robot does not need to repeat calculations or learn the same skill over and over again. This process resembles how animals learn their muscle memories in nature. The proposed method is tested on the A1 robot in a simulated environment, and it allows the robot to pronk at different speeds while precisely following the reference trajectories without heavy calculations.
Hybrid Zero Dynamics and Gait Library for Cassie Robot
We have been working on developing walking controllers for a bipedal robot Cassie built by Agility Robotics. This 3D robot has in total of twenty degrees of freedom and ten electric motors. We have been using the insights gained from the simple conservative templates to create a library of optimal gaits using full-body models implemented in an optimal control framework where the motions of every joint are taken into consideration. The current research projects include:
- Creating gait library using hybrid trajectory optimization framework C-FROST: based on the previous work at the biped robotics lab, to generate more versatile walking motions rather than moving at a constant speed, we have been building an extended set of periodic gaits that have various forward speeds, turning speeds, stride times, and terrain slopes. These solutions are optimized in parallel in a rapid gait creation framework called C-FROST where solutions are subjected to virtual constraints based on the hybrid zero dynamics.
All periodic motions are identified offline and optimized trajectories are converted to b polynomials that can be used for the online controller design.
- Stair climbing controller design with perception: the controller based on the above gait library is sufficient to reject certain amount of disturbance from the uneven terrain. However, for some specific tasks, such as stair climbing, we cannot rely solely on the controller developed for the level ground. To this end, I have been developing and testing controllers to dynamically climbing stairs with the help of LiDAR and stereo cameras.
- System identification using reinforcement learning: to overcome the obstructions imposed by the unmodelled motor dynamics and model uncertainty, we are applying data-driven methods to systematically identify the parameters in the multibody model of Cassie. This has the potential to ease the difficulty in the manual tuning of low-level controllers and speed up the implementation of specific controller design based on the gait libraries.