Boston Dynamics’ Spot has gotten a lot of coverage in the media. It is capable of performing a wide variety of tasks. With reinforcement learning and sampling-based control, Spot can stack tires autonomously. The tires in this video weigh 33lbs each, so it is an impressive feat for Spot to handle them. To pull this off, an external computer was used to perform computational operations wirelessly. An external motion capture was used for perception.
The robot autonomously coordinates manipulation and movement features to pull this off. The sampling based controller finds the best manipulation strategies to achieve a given objective. As the researchers explain: “system runs 32 parallel CPU threads, each using MuJoCo to simulate different action sequences several seconds into the future.” According to them, the tire uprighting sequence is completed in under 6 seconds per tire, which is about the same as a human.
[HT]
