Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
EMPIRIC equips robots to learn residual world models by augmenting standard physics engines with programmatic extensions that capture unmodeled phenomena like wind, adhesion, or thermal dynamics. By pairing these programmatic hypotheses with Bayesian inference over noisy observations, the agent actively designs informative experiments to resolve dynamic uncertainties and plan downstream actions. Evaluated across five simulation benchmarks and a real-world manipulation task, the approach consistently outperforms standard baselines with substantially higher sample efficiency and interpretable model transfer.
Physics engines do not need to be replaced by black-box neural networks—patching them with Bayesian-inferred code enables robots to actively discover missing physical mechanisms with drastically fewer real-world interactions.
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