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

Area of Science:

• Robotics
• Machine Learning
• Control Systems

Background:

• Robots are common in controlled settings but struggle with everyday unpredictable conditions.
• Disturbances like damage or wind gusts can cause loss of control, leading to failures.
• Existing robot controllers lack robust methods for immediate recovery from unforeseen perturbations.

Purpose of the Study:

• To develop a machine learning-based method for rapid robot control recovery.
• To enhance the resilience of robotic systems against unexpected disturbances.
• To demonstrate effective control restoration using only onboard computation.

Main Methods:

• A novel method, "Fast Learning-based Adaptation for Immediate Recovery", was developed.
• The method uses machine learning to update an onboard robot model every 225 milliseconds.
• It diagnoses and compensates for unseen perturbations to enhance existing controllers.

Main Results:

• The proposed method enabled operators to regain control and operate effectively after perturbations.
• It significantly outperformed optimal control and adaptive control baselines, which were half as effective.
• An online Deep Reinforcement Learning baseline was completely ineffective in recovering control.

Conclusions:

• Online learning enhances robotic resilience by mitigating perturbation impacts.
• The fast adaptation method allows robots to maintain operability in dynamic, real-world scenarios.
• This approach is crucial for integrating robots into everyday environments.