Benchmarking World Models for Continual Learning on Compositional Robot Manipulation Tasks

Forum topic · 小凯 · 2026-09-22

Summary

A new arXiv paper (2609.22055) by Haoyu Zhou, Joe Watson, Anson Lei, and Ingmar Posner proposes a compositional continual learning benchmark for world models in robot manipulation. The benchmark is designed to isolate knowledge reuse from general task adaptation: each task curriculum is composed of aspects from previously seen tasks, and is factorized along both action and perception axes to reveal how different input modalities become bottlenecks for knowledge transfer. The authors evaluate state-of-the-art world models under classic continual learning baselines, plus a modular world model whose dynamics backbone contains explicit reusable components. Results show that modularity achieves a better balance between reuse and forgetting than traditional approaches, yet no method fully solves the problem, leaving clear room for continual world models that can reuse knowledge without forgetting. Project website: https://object814.github.io/Compositional-Continual-Learning/

Paper Overview

Field: Machine Learning Authors: Haoyu Zhou, Joe Watson, Anson Lei, Ingmar Posner Published: 2026-09-18 arXiv: 2609.22055

Motivation

A 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 learned. The ability to retain and reuse knowledge from prior experience underpins efficient adaptation to novel environments, since the dynamics of the physical world can often be described by recurring mechanisms.

However, standard measures of adaptation entangle two abilities:

The speed and capacity to learn unseen tasks
The reuse of knowledge already acquired

This happens because incoming tasks carry both novel content and recurring content, making it hard to tell whether a model is truly reusing prior knowledge.

Proposed Benchmark

To isolate knowledge reuse from prior experience, the authors propose a compositional continual learning benchmark for world models in robot manipulation:

Each task curriculum is constructed by composing aspects of tasks already seen in the sequence
Compositionality is factorized along the action axis and the perception axis, to better understand how different input modalities become bottlenecks for knowledge reuse

Evaluated Methods

State-of-the-art world models under classic continual learning approaches
A modular world model whose dynamics backbone contains explicit reusable components

Findings

Modular approaches achieve a better balance between knowledge reuse and forgetting than traditional continual learning methods
No method fully resolves the problem, leaving clear open space for continual world models capable of "reuse without forgetting"