EMPIRIC: Experiment-Driven Learning of Residual World Models for Robot Planning

2026-09-28

AI Summary

This study addresses the challenge of robots interacting with objects governed by mechanisms unmodeled in standard physics engines. To this end, it proposes an interpretable and reusable residual world model that leverages program synthesis to generate code extending the physics engine, thereby compensating for missing dynamics. By integrating Bayesian inference with active experimental design, the framework achieves data-efficient mechanism learning and motion planning. In simulation, it efficiently completes complex tasks, while on a real robot, it successfully infers latent physical parameters such as wind disturbances and domino masses. These capabilities enable precise prediction and manipulation control, demonstrating the framework’s effectiveness in bridging the gap between simulated and real-world physical interactions.

Abstract

A robot should be able to learn through experiments how unfamiliar objects behave and interact, then plan with that knowledge. It need not start from scratch: physics engines supply knowledge of motion and contact, but can omit entire mechanisms, such as glue curing, water heating, or wind. We present EMPIRIC, an agent that learns a residual world model: a physics engine extended with code for the missing mechanisms. The learned programs can introduce new forces, constraints, and hidden state, and Bayesian inference estimates their parameters and states from noisy observations. The resulting model lets the agent predict the outcomes of actions, choose informative experiments, and revise its hypotheses when predictions fail. Across five simulated domains, EMPIRIC learns interpretable, reusable models, and solves more tasks with fewer environment interactions than all three baselines. On a physical robot, it learns wind forces and domino masses to solve a manipulation task. Website and code: https://yichao-liang.github.io/empiric

Problem

• Robot Planning
• World Models
• Physics Engines
• Experiment-Driven Learning
• Missing Mechanisms

Innovation

• Residual World Models
• Experiment-Driven Learning
• Bayesian Inference
• Robot Planning
• Program Synthesis