Researchers led by Luzia Knoedler have built Sim2CBF, a safety filter that keeps a robot from failing using only forward predictions of what its actions will do, according to a paper on arXiv. Many useful predictive models, such as physics simulators or learned world models, do not give the reliable derivatives that standard control barrier function filters need.

Sim2CBF approximates the barrier condition with finite differences and local interpolation over sampled commands, while keeping its guarantees against failure. The barrier itself comes from rollouts of a backup policy. It needs fewer samples than purely sampling-based filters and holds the robot back less than simply blending in the backup policy.

The team tested it in simulation, on a real robot toppling boxes and in control from camera images, showing it works with contact-rich and learned dynamics and adapts when the predictive model changes. The abstract gives no figures.

Source: arXiv Robotics (cs.RO)

Background

Luzia Knoedler

Robotics researcher

Robotics researcher.