Getting robots back on track by reconstituting control in unexpected situations with online learning

Robotic systems are increasingly common in strictly controlled environments (e.g. warehouses), but they have yet to fully integrate into our everyday lives. Everyday scenarios are a challenge for robot controllers due to disturbances that can be caused by unforeseen conditions (e.g., damage, flat tires, or wind gusts) or unexpected usage. In these cases, operators may lose control of their systems, leading to potentially catastrophic outcomes, such as crashes or failed operations. Our method, "Fast Learning-based Adaptation for Immediate Recovery", uses machine learning to counter loss of control by enhancing existing controllers. It rapidly diagnoses and compensates for unseen perturbations by updating an onboard model every 225 milliseconds. Even in a setting with only onboard compute, we show that operators equipped with our method regain control and operate as effectively as in unperturbed conditions across a wide range of perturbations. Other state-of-the-art methods, such as an optimal control and an adaptive control baseline, were found to be half as effective at recovering from perturbations, while an online Deep Reinforcement Learning baseline proved entirely ineffective. In this work, we demonstrate that our online learning method enhances robotic resilience by mitigating the impact of perturbations on system operability.