Researchers led by Thomas A. Berrueta, with Soon-Jo Chung among the authors, have developed recoverability, a way to give formal safety guarantees to robots running behaviours that cannot be certified directly, such as learned policies and aggressive manoeuvres at the limits of traction, according to a paper on arXiv. Usual safety certificates need accurate models at the edge of the operating range, where models are least reliable.

Rather than proving that the deployed behaviour stays safe, the method certifies that the robot can always re-engage a trusted recovery mode. The authors prove that switching to that mode from anywhere inside a certified tube leads to exponential convergence despite disturbances and model errors, and they work out when to switch.

In simulated quadrotor flights through uncertain obstacles it beat safety filters based on invariance. On a full-scale IndyNXT autonomous race car it repeatedly brought the car back from an aggressive racing line at the limits of handling in the rain.

Source: arXiv Robotics (cs.RO)

Background

Thomas A. Berrueta

Robotics researcher

Robotics researcher.