Researchers led by Samuel Bateman introduced Action Influence Maximization (AIM), an algorithm that actively gathers training episodes to make video‑based world models pay attention to commanded actions. In tests on simulated navigation and pixel‑based manipulation, AIM added just 500 new episodes to an existing set of 3,000 expert demos and matched the controllability of a model trained on 300,000 expert episodes. The method outperformed standard exploration and intrinsic‑motivation techniques in data efficiency.
Why it matters
For robot developers, AIM means they can train more responsive control models with far fewer costly demonstration recordings.
