The reports autonomous rolling, uprighting and dragging of a 15 kg tire using onboard depth cameras and computing. The tire is heavier than the robot arm can lift on its own, so the system coordinates its arm, legs and body.
Earlier work depended on ceiling-mounted motion-capture cameras, a desktop workstation and a script to switch between skills. The new system uses camera-based policies and a gating network that chooses the next skill on the robot.
The demonstration includes recovering after a person knocks the tire over. Testing used one tire on flat ground. The results do not establish that the same performance extends to other objects or uneven terrain.
Source: rai-inst.com
Limits
One tire on flat ground. Results on other objects and terrain remain open.