Title: Identifying the Conditions Under Which Predictive Inverse Dynamics Surpasses Behavior Cloning

Abstract:

Behavior cloning (BC) stands as a widely adopted offline imitation learning technique; however, its effectiveness frequently diminishes when expert demonstration datasets are scarce. To address this limitation, recent research has proposed predictive inverse dynamics models (PIDM), an architectural framework that integrates an inverse dynamics model with a predictor for future states. Although empirical evidence suggests that PIDM typically yields superior results compared to BC, the underlying mechanisms driving this advantage have lacked rigorous theoretical clarification.

In this study, we offer a theoretical rationale, positing that PIDM operates by navigating a bias-variance tradeoff. Specifically, while the act of forecasting future states introduces a degree of bias, conditioning the inverse dynamics model on these predictions can substantially mitigate variance. We derive specific conditions regarding the bias of the state predictor that must be met for PIDM to achieve both lower prediction error and improved sample efficiency relative to BC. Furthermore, we demonstrate that the performance gap between the two methods expands as additional data sources become accessible.

Our theoretical findings are substantiated through empirical validation across diverse environments. In 2D navigation tasks, BC necessitates up to five times more demonstrations—and an average of three times more—to match the performance levels of PIDM. Similarly, in a complex 3D setting within a modern video game characterized by high-dimensional visual inputs and stochastic transitions, BC requires more than 66% additional samples to achieve comparable outcomes to PIDM.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC