Shared Autonomy via Deep Reinforcement Learning

Siddharth Reddy, Anca D. Dragan, Sergey Levine

University of California, Berkeley

Links to pre-print, code, and blog post

Our goal is to assist humans with real-time control tasks. We propose a deep reinforcement learning framework for shared autonomy which combines a human pilot with a robotic copilot that adapts to the user.

In one of our experiments, we assisted human pilots with "perching" a quadrotor: navigating to a small landing pad and pointing the drone's camera at a random object.

human-solo-quad (1).mp4

Human Pilot (Solo)

The pilot's display only shows the drone's first-person view, so pointing the camera is easy but finding the landing pad is hard.

human-assisted-quad (1).mp4

Human Pilot + RL Copilot

The copilot doesn't know where the pilot wants to point the camera, but it knows where the landing pad is. Together, the pilot and copilot succeed at the task.

In another experiment, we helped human pilots play the Lunar Lander game.

human-solo-lander (1).mp4

Human Pilot (Solo)

human-assisted-lander (1).mp4

Human Pilot + RL Copilot

Humans rarely beat the Lunar Lander game on their own, but with a copilot they do much better.

We asked users for their subjective evaluations after each experiment:

rss18-subjective-evals.pdf