RoboticsRobot simulationStory 04
Action prediction keeps a world model focused on physics
What changedConcat IDM trains the shared encoder to reconstruct actions from consecutive latent states, without changing the test-time planner. Across three Push-T runs with 24 appearances, planning success rose from 9.1 to 77.2 percent, isolating a failure that visual variation can cause.
The useful part
Why it matters
Across three Push-T runs with 24 appearances, planning success rose from 9.1 to 77.2 percent, isolating a failure that visual variation can cause.
Keep in mind
Good to know
The evaluation is controlled and simulation-based.
Evidence