Learning A Physical Long-term Predictor
This provides a novel approach for AI systems to make actionable long-term predictions in physical domains without explicit modeling of physics, which could benefit robotics and autonomous systems.
The paper tackles the problem of long-term prediction of mechanical phenomena from sensor data by using a single neural network for end-to-end prediction, showing it can outperform methods with ground-truth simulators, especially when physical parameters are unobserved or unknown.
Evolution has resulted in highly developed abilities in many natural intelligences to quickly and accurately predict mechanical phenomena. Humans have successfully developed laws of physics to abstract and model such mechanical phenomena. In the context of artificial intelligence, a recent line of work has focused on estimating physical parameters based on sensory data and use them in physical simulators to make long-term predictions. In contrast, we investigate the effectiveness of a single neural network for end-to-end long-term prediction of mechanical phenomena. Based on extensive evaluation, we demonstrate that such networks can outperform alternate approaches having even access to ground-truth physical simulators, especially when some physical parameters are unobserved or not known a-priori. Further, our network outputs a distribution of outcomes to capture the inherent uncertainty in the data. Our approach demonstrates for the first time the possibility of making actionable long-term predictions from sensor data without requiring to explicitly model the underlying physical laws.