Researchers develop 'counterfactual' world models to help robots learn from surprises
New technique could help robots infer hidden factors from unexpected outcomes and improve future predictions and decisions
Researchers from Brown University, New York University and AMI Labs have developed a technique for 'counterfactual world models' that could enable robots to learn from unexpected events by inferring hidden factors, such as currents or loads, from the gap between prediction and reality. The approach, described in a pre-print paper, could help robots adapt to surprises in real-world settings, such as handling miscalibrated motors, without needing to be trained for every possible scenario.