Habitat-Lab Quadruped Embodied AI Research

Abstract

With the rise of reinforcement learning, a number of physics engines and frameworks have been used to simulate virtual environments. Additionally, various benchmarks have been popularized in order to assess the ability of RL algorithms to learn from these environments and produce effective policies for agents to act on. In this project, we aim to replicate a popular reinforcement learning benchmark, Open AI's Ant-v2 environment, in the AI Habitat platform and apply Proximal Policy Optimization to train a quadruped robot to run on a platform. Doing so allows us to extend the standard benchmark by adding visual sensors to the robot and including more complex environments such as indoor spaces. This fits into a larger goal of solving point-navigation tasks with dynamic locomotion.