New iKCE Metric: World Models Skip Physics, Not Motion

A workshop paper accepted at the RSS Robot World Model Workshop 2026 makes an argument worth chewing on. The standard story about learned world models failing at long horizons, generic 'compounding error', is actually two different failures glued together, and only one of them is the interesting one.

The paper, Imagined Rollouts are Kinematic, Not Dynamic, from Finn Rasmus Schäfer, Korbinian Moller, Yuan Gao, Christian Oefinger, Sebastian Schmidt and Johannes Betz, introduces a diagnostic they call the imagined Kinematic-Consistency Error, or iKCE. The idea is straightforward. Measure, per step, how far an imagined rollout departs from a closed-form kinematic null, the trajectory you'd get from position and motion alone with no forces required. If a model is doing kinematic bookkeeping well, iKCE stays low. If it also captures the underlying dynamics, then when the dynamics change the iKCE signal should move.

Their headline number is on a DreamerV3 checkpoint trained on the DMC walker-walk task. Imagined iKCE runs roughly two orders of magnitude above matched real-physics rollouts. That is not a small drift. And in a friction sweep that crosses a gait-collapse boundary, the point where the physical policy stops working because the ground got too slippery, the model's iKCE stayed statistically flat even as the trained policy's reward collapsed. The world model, in other words, kept imagining the walker walking through the transition. It did not feel the friction change at all. The authors are careful to scope the claim to horizons longer than the embodiment's gait period, which is where the split shows up.

Take the specifics as reported, not settled. This is a nine-page workshop paper, tested on one Dreamer checkpoint on one task, and the honest caveat is that iKCE's value on contact-rich manipulation, deformables, or transformer and diffusion world models is not shown here. The reporting also doesn't tell you which training interventions actually close the gap the diagnostic reveals.

But the framing is the part worth watching. If you are building on Dreamer-style planners for robotics, the practical read is to stop treating kinematic plausibility and dynamic correctness as one evaluation axis. A rollout that looks physical and a rollout that respects physics are not the same thing, and this paper gives you a cheap way to tell them apart.