Rather than assessing an approach on the basis of a brief trial, the method predicts its long-term effectiveness. Across 108 simulated scenarios, it reduces the number of failed recoveries by 27.8 points.
When a robot with legs is damaged, the recovery algorithms generally test several approaches over short tests and retain the best one. The problem is that a gait that is effective for a few seconds can then degrade, and without a reset of the robot, each test modifies the physical state that the next one inherits.
The PGRR (persistent profile-guided reset-free recovery) method predicts several performance indicators on a prolonged execution of each approach and relies on conservative estimates of uncertainty to orient brief in-line probes. Evaluated on 108 tasks simulated under MuJoCo covering leg amputation, distal fracture and joint blockage, it improves the final quality of the chosen gait compared to a reference to a Gaussian process inspired by IT&E, and reduces low-quality failures by 27.8 percentage points. Gains appear in all three types of damage; ablations point to conservative uncertainty management as the main factor.
Read the full article on the original source
Go to the source →