Adapt On-the-Go: Behavior Modulation for Single-Life Robot Deployment

Videos of Go1 Real-World Trials

Task 1: Heavy Luggage

Task 2: Dynamic Load

Task 3: Roller Skates

Sample Rollout in Simulation

In our simulated experiments, we find that ROAM is over 2x as efficient as prior methods that are designed for fast adaptation. Here we plot the behavior distribution for ROAM over the course of a single-life trial, where the agent is tasked with adapting to different stiffness on-the-go. Green bars indicate relevant behaviors to the current situations while red bars indicate irrelevant behaviors. We find that ROAM can quickly react to changing situations by choosing and adapting relevant behaviors on-the-fly.