Quadrupedal Locomotion via Central Pattern Generator(CPG) and Reinforcement Learning(RL)
Overview
In this project, we control the locomotion of a quadrupedal robot via CPG and RL respectively.
The CPG controller generates different gaits according to predefined parameter sets and keeps balance after tuning, which . The maximum speed reached with the gait pronk, whereas the fastest gait for real dogs is rotatory gallop. In our opinion,
- for a faster locomotion with simple CPG, more sensors are needed to keep balance and seize the right moment to swing. (This implies the importance of reflex)
- CPG provides a convenient and explainable approach to convert the gait to specific physical parameters
- however, while it doesn’t take much effort to tune the parameters for most gaits, it’s a hard to make the robot move at a speed comparable to a real dog for some gaits.
We used 2 reward functions of different types and used both Proximal Policy Optimization (PPO) and Soft Actor Critic (SAC) in our experiments. The quadrupedal robot achieves a fastest speed at 4.8m/s
Some Result Demos
- Slow walking using CPG
- Quadruped galloping gait in the competiton environment with locomotion controller trained by DRL (PPO).
- Quadruped walking gait in the plain environment trained by DRL (SAC).
- Quadruped back walking in the competition environment trained by DRL (SAC) with joint PD control method.
For more info
comments powered by Disqus