In this project, we will use a quadruped robot and train it to perform navigation on long timescales through a human imitation approach. Legged robots are currently capable of learning low-level skills such as walking and recovering from falling, using reinforcement learning algorithms in a fast simulator. However, learning higher-level skills such as navigation remains elusive due to two main challenges: First, navigation takes place over long timescales, which makes it challenging for reinforcement learning algorithms to explore and attribute reward. Second, navigation relies more on vision at a distance rather than proprioception, which raises an issue since the sim-to-real gap is larger for vision than proprioception.

To address this gap, we will design a system for learning navigation from real-world human demonstrations. First, to create a controlled testbed for our method, we will procedurally generate difficult terrain in simulation, extending existing research. Next, we will manually collect data to provide examples of good behavior in real life. Using the simulation and manually collected examples, we will train a navigation policy that will allow the quadruped to travel point-to-point while prioritizing safety locally. As a final extension, we will attempt to implement navigation from a third-person perspective holding a camera from behind that’ s facing the robot while navigating. This can help to overcome the engineering challenge and expense of working with cameras mounted onboard the robot.

I’m participating in SuperUROP because I believe it will be a great opportunity to gain advanced research experience. Through machine learning classes and previous UROPs, I developed an interest for working on machine learning problems in a research setting. I hope to make meaningful contributions towards the intersection of neuroscience and machine learning and to gain the skills necessary to lead my own research projects in the future.