Event
WorldEcho Evaluates Action-Following in World Models
1 reports · 1 sources
Sixiang Chen and other authors propose an action-conditioned world model evaluation method, WorldEcho, and an alignment method, WorldSync. WorldEcho, through visual integrity and SE(3) end-effector trajectory alignment, tests whether models truly follow instructed actions across a broader action distribution. The authors state that diagnostics show models trained on expert data perform relatively well on demonstration actions, but on non-expert trajectories produce action-inconsistent or visually invalid generations. The authors state that experiments show WorldSync reduces integrity-gating error on RoboTwin; when used for iterative policy improvement, it achieves higher success rates in simulation and on real robots.
Report timeline
- Do Robotic World Models Really Follow Actions? WorldEcho Diagnoses and WorldSync Aligns Action-Conditioned Generation
The authors propose WorldEcho, which evaluates whether action-conditioned world models truly follow commanded actions over a broader action distribution through visual integrity and SE(3) end-effector trajectory alignment. The diagnosis shows that models trained on expert data perform well on demonstrated actions but produce action-inconsistent or visually invalid generations on non-expert trajectories.
arXiv cs.RO ·