When Does Predictive Inverse Dynamics Outperform Behavior Cloning?
A new research paper explores the effectiveness of Predictive Inverse Dynamics Models (PIDM) compared to Behavior Cloning (BC) for imitation learning. The study theoretically explains that PIDM offers a bias-variance tradeoff, where predicting future states introduces bias but reduces variance, leading to better performance and sample efficiency than BC, especially with limited expert demonstrations. Empirical tests in navigation and video game environments confirmed that PIDM requires significantly fewer samples to achieve comparable results to BC. AI
IMPACT Explains why PIDM is more sample-efficient than behavior cloning, potentially guiding future imitation learning applications.