Movendi Research Note No. 1

Toward Predictive Models of Reality

Abstract

Movendi is developing Reality Models — AI systems that learn to predict the evolution of the physical world from observation. This note introduces the Direct Predictive Model (DPM), a class of self-supervised models intended to learn the dynamics of physical reality directly from experience, and sets out the hypothesis that intelligence emerges from prediction.

Foreword

Movendi Research Notes document the evolution of our research toward predictive models of reality. They are intentionally published at an early stage to encourage scientific discussion and collaboration with the broader research community.

Rather than presenting finished conclusions, these notes describe the hypotheses, architectural choices, experiments, and observations that shape our work. As our understanding evolves, these ideas will continue to mature through experimentation, collaboration, and constructive criticism. We welcome discussion with researchers, students, and institutions who share an interest in advancing predictive intelligence.

Executive Summary

Artificial intelligence has achieved remarkable progress in language, vision, and generative modeling. More recently, significant research efforts have begun exploring predictive representations of physical environments through so-called world models, in which learned latent dynamics provide an internal representation for prediction, planning, or control (Ha and Schmidhuber, 2018; Hafner et al., 2025).

Over the past three years, Movendi has approached this challenge from a different direction. Through the deployment of AI systems in manufacturing and healthcare, we have developed operational representations of reality by combining digital twins, computer vision, telemetry, state machines, control systems, machine learning models, and large language models. These systems solve real-world problems under uncertainty and partial observability, but they require increasingly complex engineering and integration.

This experience has led us to a broader scientific question.

Can a machine learn these representations directly from experience, rather than through engineered combinations of specialized models and software components?

Movendi was founded around a simple research hypothesis: intelligence emerges from prediction — not the prediction of words, images, or pixels, but the continuous prediction of the evolution of observed reality.

To explore this hypothesis, we are developing the Direct Predictive Model (DPM): a new class of self-supervised models designed to learn the dynamics of physical reality directly from experience. Our long-term objective is to understand whether concepts such as object permanence, causality, support, collision, and physical reasoning can emerge naturally from predictive learning, without explicit physics engines, symbolic rules, or synthetic simulation. Prior research has demonstrated that predictive models can acquire measurable expectations concerning object permanence, continuity, and solidity, although under substantially more structured conditions (Piloto et al., 2022).

Why Now?

Recent advances in self-supervised and representation learning have shown that useful abstract representations can be acquired by predicting missing information in representation space, without reconstructing every detail of the sensory input (Assran et al., 2023). At the same time, our experience deploying AI in industrial and healthcare environments has highlighted the practical limitations of systems assembled from multiple specialized models connected through increasingly sophisticated engineering.

We believe the next generation of AI systems should learn these predictive representations directly, progressively replacing engineered integrations with unified predictive models capable of maintaining an evolving estimate of reality.

References

1. David Ha and Jürgen Schmidhuber, “Recurrent World Models Facilitate Policy Evolution,” Advances in Neural Information Processing Systems 31, 2018, pp. 2450–2462. Link
2. Mahmoud Assran et al., “Self-Supervised Learning From Images With a Joint-Embedding Predictive Architecture,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 15619–15629. Link
3. Luis S. Piloto, Ari Weinstein, Peter Battaglia, and Matthew Botvinick, “Intuitive Physics Learning in a Deep-Learning Model Inspired by Developmental Psychology,” Nature Human Behaviour, vol. 6, 2022, pp. 1257–1267. Link
4. Danijar Hafner et al., “Mastering Diverse Control Tasks Through World Models,” Nature, vol. 640, 2025, pp. 647–653. Link