For a robot to navigate new environments, it needs to be able to “see” what’s around it. By replacing external motion capture with a three-stage perception pipeline and reinforcement learning, fully autonomous, untethered robotic manipulation was achieved in an unmapped environment - in this case, a nearby skatepark. As privileged state inputs are progressively distilled into noisy depth-camera observations, the policy adapts to handle intermittent visibility and sensor corruption, resulting in a risk-averse yet highly dynamic execution of complex tasks like tire rolling and re-orientation.
Learn more about how we built and trained this on board system
Learn more about how we built and trained this on board system
This website uses cookies
Comments 0
Be the first to comment!