Haddy Alchaer reports planning methods that lift JEPA world model success

Haddy Alchaer published findings from 600 experiments on LeWM, a JEPA world model released this year by a team including Yann LeCun and benchmarked on the Push-T robot pushing simulation. On his own site he reports that, when given expert subgoals, substituting the simulator's exact physics for LeWM's learned predictor moved success by one point, 78% versus 77%, while chaining 25-step subgoals raised success on 75-step goals from 22% to 80% without retraining. Adding a cost to jumps between demonstrations in a graph of embedded expert frames lifted a memory-based planner from 62% to 85%. He also argues the standard Push-T test is easier than it looks, since its demonstrations all end in about the same place, and tries a harder variant that draws start and goal from different demonstrations.