Event
CoRL 2026 Paper Proposes AIM Algorithm to Improve World Model Controllability
1 reports · 1 sources
Samuel M. Bateman and other researchers propose the Action Impact Maximization (AIM) algorithm in a CoRL 2026 paper, which actively collects training data to increase the action impact of the dataset, thereby improving the controllability of action-conditioned video world models. The authors prove that a data's action impact metric upper-bounds the action impact any model trained on that data can express, and that the narrow action distribution of expert demonstrations causes this upper bound to be low. On simulated navigation and manipulation tasks, AIM is compared with exploration heuristics and intrinsic motivation methods: in pixel-based manipulation tasks, AIM is the only method that consistently improves controllability; on end-effector controllability, adding 500 actively collected clips to 3,000 demonstrations can achieve the effect of a world model trained on 300,000 expert clips, reducing data volume by 85 times.
Report timeline
- Maximizing Action Influence to Train Controllable World Models
A CoRL 2026 paper proposes the Action Influence Maximization (AIM) algorithm, which actively collects training data to increase the action influence of the dataset, thereby improving the controllability of action-conditioned video world models. The authors prove that a data's action influence metric upper-bounds the action influence any model trained on that data can express, and that the narrow action distribution of expert demonstrations causes this ceiling to be low.
arXiv cs.RO ·