Continual world model with counterfactual simulated interactions

This paper proposes a continual world model that improves robustness by prospectively simulating the consequences of a model's own predictions and rehearsing on those counterfactual trajectories. The setting is bionic limb behavior modeling, where a predictor $f$ maps recent residual-body sensor histories to desired prosthetic joint profiles across multiple locomotion tasks (for example, level walking, stairs, ramps). The core contribution, called multitask prospective rehearsal (MPR), couples $f$ with a learned prospective dynamics model $g$ that forecasts the next sensor state given the current state and an action (the predicted joint profile). By rolling $g$ forward on $f$'s own outputs to generate counterfactual inputs (the predicted next sensor state) and pairing these with the future target joint profile, the method creates error-aware rehearsal data that explicitly conditions $f$ to correct the distribution shift it induces. This enables continual, multitask adaptation without online environment interaction or access to expert labels at deployment.