Planning changes raise LeWM simulation success from 22% to 85%

A researcher reports raising LeWM's success rate on long-horizon simulated robot tasks from 22% to 85% without retraining the model. The experiments used Push-T, in which a robot pushes a T-shaped block into a target position, with goals 75 steps away. The method divides planning into shorter segments and selects intermediate goals from a graph of about 460,000 embedded states from training demonstrations. Penalizing jumps between demonstrations helped avoid physically implausible paths. The 85% result covered the same 250 long-horizon tasks and matched a planner given intermediate goals from expert demonstrations; the tests were conducted in simulation.