Benchmarking World Models for Continual Learning on Compositional Tasks
A new compositional benchmark finds modular world models reuse robot skills better, but still forget.
The paper proposes a compositional continual-learning benchmark for world models in robot manipulation, designing task curricula that recombine previously seen action and perception factors. It evaluates state-of-the-art world models with canonical continual-learning methods and a modular model whose dynamics backbone has explicitly reusable components. Modularity balances reuse against forgetting better than conventional methods, but none fully solves continual adaptation without forgetting.
- Benchmark uses compositional robot-manipulation curricula to isolate knowledge reuse.
- Tasks are factorized along action and perception to locate reuse bottlenecks.
- A modular world model reuses dynamics components better than standard continual-learning methods.
- No evaluated method fully reuses prior knowledge without forgetting.
From the original article
1 min read at arxiv.orgA desirable property of a world model is the ability to learn continually across tasks, adapting to new environments without forgetting what the agent has already learnt. In particular, the ability to retain and reuse knowledge obtained from prior experiences underpins an agent's ability to efficiently adapt to novel environments, as the dynamics of the physical world can often be described in recurring mechanisms.
By Haoyu Zhou · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.