← Research

Movendi Research Note No. 3

Persistence is a correspondence problem

A new scan does not observe the same points again.

Sensors resample visible surfaces from changing positions and angles. Some regions disappear behind occlusions. Others become visible for the first time. Reflective, deformable, granular, or moving materials may produce substantially different measurements from one observation to the next.

The perception problem is therefore not simply to match every point in one scan with a point in the next. In many cases, no exact one-to-one correspondence exists.

A Reality Model must distinguish between two kinds of points.

Observation points are temporary measurements produced by a sensor.

Model points belong to the model’s persistent estimate of the environment.

This distinction connects to the broader requirement for persistent, revisable physical state developed by Leroy & Leroy (2026) [3]. Their position paper leaves the representation open; the point-based formulation explored here is one candidate implementation.

At each moment, the model predicts where its existing points - and the relationships among them, should be next. The new scan is then compared with that predicted state. Measurements may update existing model points, support several possible correspondences, reveal previously occluded regions, or provide evidence that the model’s earlier estimate was wrong.

Some model points may remain present while unobserved. Some may divide or merge as the model changes its spatial resolution. New points may be introduced when new physical structure becomes visible. Others may lose confidence and eventually disappear from the representation.

Correspondence should therefore be probabilistic rather than forced.

It may depend on position, motion, appearance, local geometry, neighboring points, previous observations, and the actions occurring in the environment.

When individual correspondence is ambiguous, the model may preserve identity at the level of a neighborhood, material region, or moving structure rather than assigning an unjustified point-to-point match.

Object-centric video models provide a related precedent for representing and tracking entities across frames (Kipf et al., 2022) [5].

This turns correspondence into a predictive problem.

Prediction already supports control through learned world models (Hafner et al., 2025; Wu et al., 2023) [1], [6], action-conditioned visual simulation (Yang et al., 2024) [2], and visual representation learning (Bardes et al., 2024) [4]. These results motivate prediction as a learning signal. The hypothesis here is that prediction can also help establish physical correspondence across changing observations.

The question is not only:

Which point in the current scan matches this point in the previous scan?

It is:

Given everything the model believed about the environment, where should this physical structure now be - and which current measurements provide evidence about it?

Prediction proposes continuity. Perception corrects it.

A successful Reality Model would maintain physical persistence through resampling, motion, deformation, occlusion, disappearance, and reappearance. The ability to establish this continuity may be one of the first measurable forms of physical intelligence to emerge from prediction.

Citations revised 26 October 2026.

References

  1. Hafner, D., Pasukonis, J., Ba, J., and Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature, 640, 647–653.
  2. Yang, S., et al. (2024). Learning Interactive Real-World Simulators. International Conference on Learning Representations (ICLR).
  3. Leroy, L. H., and Leroy, M. (2026). Position: World Models of Physical Sites Must Discover Their Own State. World Models in Physical AI Workshop, NeurIPS 2026. Workshop poster.
  4. Bardes, A., et al. (2024). Revisiting Feature Prediction for Learning Visual Representations from Video. Transactions on Machine Learning Research (TMLR).
  5. Kipf, T., et al. (2022). Conditional Object-Centric Learning from Video. International Conference on Learning Representations (ICLR).
  6. Wu, P., Escontrela, A., Hafner, D., Abbeel, P., and Goldberg, K. (2023). DayDreamer: World Models for Physical Robot Learning. Proceedings of the 6th Conference on Robot Learning, PMLR 205, 2226–2240.