Multi-Goal RL for Legged Robot Footstep Planning
Iterative optimization method using multi-goal reinforcement learning for real-time dynamic footstep planning over a receding horizon for legged robots.
To address the problem of dynamic footstep planning for legged robots, we developed an iterative optimization method that plans in real-time over a receding horizon. We provided an analytical solution to the linearization of event-based discrete hybrid dynamics, allowing superior accuracy with lower computation and time complexity compared to numerical differentiation.
We also established a new method to incorporate terrain as a soft-state constraint into the iLQR formulation.
Foothold optimization conforming to terrain cost map using gradient descent of local polynomials.
Tags: Motion Planning · Controls · Deep Learning · Reinforcement Learning · Legged Robots