Trajectory based deep policy search for quadrupedal walking

Published in RO-MAN, 2019

Recommended citation: S Kolathaya, A Joglekar, S Shetty, D Dholakiya, A Sagi, S Bhattacharya, A Singla, S Bhatnagar, A Ghosal, B Amrutur, Trajectory based deep policy search for quadrupedal walking, 28th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN) https://ieeexplore.ieee.org/document/8956369

In this paper, we explore a specific form of deep reinforcement learning (D-RL) technique for quadrupedal walking—trajectory based policy search via deep policy networks. Existing approaches determine optimal policies for each time step, whereas we propose to determine an optimal policy for each walking step. We justify our approach based on the fact that animals including humans use “low” dimensional trajectories at the joint level to realize walking. We will construct these trajectories by using Bezier polynomials, with the coefficients being determined by a parameterized policy. In order to maintain smoothness of the trajectories during step transitions, hybrid invariance conditions are also applied. The action is computed at the beginning of every step, and a linear PD control law is applied to track at the individual joints. After each step, reward is computed, which is then used to update the new policy parameters for the next step. After learning an optimal policy, i.e., an optimal walking gait for each step, we then successfully play them in a custom built quadruped robot, Stoch 2, thereby validating our approach.

Download paper here