An Oxford benchmark has found that a modular world model can preserve old robot skills far better than a conventional model, yet still fails to learn a final recombined skill any faster. The six-suite simulation study separates retention from reuse, showing that adding reusable dynamics modules addresses only one side of continual learning.
A world model predicts how an environment will change after an action, giving a robot an internal simulator for planning. Haoyu Zhou and colleagues at the University of Oxford built curricula in which a robot first learned primitive tasks and then faced a task made from those earlier parts. Two suites recombined actions, two changed visual conditions, and two changed both action and perception.
The team adapted DreamerV3, TD-MPC2 and a modular Prismatic World Model, or PWM, to the same Meta-World manipulation setup. Each observation combined three camera views with the robot arm's pose, velocity and gripper state. Backward transfer measured how much earlier tasks deteriorated, while forward transfer measured how quickly the final composition task was learned relative to training it from scratch. Scores were normalized between a failure floor of 0 and a scripted-policy reference of 100, with results averaged over three random seeds.
Standard continual-learning methods exposed a trade-off rather than a solution. Experience replay produced the strongest forward transfer but kept 5% of every previous task's transitions, so its memory use grew with the curriculum and forgetting only slowed. PackNet pushed forgetting close to zero by freezing task-specific subnetworks, but its pool of unused parameters shrank after every task and limited how long the sequence could become.
PWM took a different route. It added three dynamics experts for every incoming task, froze older experts and trained a router to combine all available modules. To isolate the effect of that modular dynamics design, the researchers also pretrained an image encoder on demonstrations from every benchmark task and then froze it. They gave the same encoder checkpoint to TD-MPC2 for the direct comparison.
Across the six suites, PWM recorded an average backward-transfer score of 3.85, compared with 34.97 for TD-MPC2, where lower is better. Its forward-transfer score was 36.18 against 38.11 for TD-MPC2, however, so the modular model did not improve the speed of reuse. This is an internal comparison under the paper's privileged frozen-encoder condition rather than a claim that PWM sets a new general benchmark.
The failure pattern also matters. Forgetting worsened as the tests moved from action changes to visual changes and then to full composition. PWM's router recovered the expected reuse pattern in several curricula, and the expert favored during a primitive task stayed top-ranked after composition learning in 13 of 14 cases. But the router chose one allocation for an entire task, which prevented it from switching experts as multi-stage tasks changed phase.
The authors therefore treat the result as evidence about architecture, not a deployment result. The experiments used simulated tasks, one model size and one expert count, without controlled parameter or compute budgets. The encoder also had advance access to demonstrations from the full task set; without that privilege, PWM performed about the same as TD-MPC2.