Learning-based Control

Legged robots often face challenges due to their high degrees of freedom and the difficulty of accurately modeling complex environments. These factors discourage the development of robust and stable controllers. However, by leveraging reinforcement learning(RL) techniques, we have successfully developed locomotion controllers that enable the quadrupedal robots to adapt and traverse through challenging terrains such as rough, slippery, and deformable surfaces. This learning-based approach not only improves robustness and stability but also reduces dependency on accruate and computationally intensive models, making our controllers well-suited for practical applications in dynamic and unpredictable environments.

../_images/VariousTerrains.png

Learning-based locomotion on various terrains

Through extensive experimentation, we have demonstrated the effectiveness of our RL-based locomotion controllers. The list of our work is shown below.

  • Suyoung Choi et al., Learning quadrupedal locomotion on deformable terrain. Sci. Robot. 8, eade2256(2023) paper video

../_images/DeformableContactModel.png

Contact model of deformable terrains

  • Gwanghyeon Ji et al., “Concurrent Training of a Control Policy and a State Estimator for Dynamic and Robust Legged Locomotion,” in IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 4630-4637, April 2022 paper video

../_images/ConcurrentControlDiagram.png

Concurrent training architeture

  • Blind locomotion through the natural environment characterized by steep slopes and scattered debris (ongoing) video1 video2

  • Whole body control for pushing objects (ongoing)

  • Dynamic locomotion for traversing stepping stones (ongoing)