Andrii Bidochko

• Updated: August 24, 2026
• 3 min read

Towards Human Motion World Models via Executable Behaviour Representations – A Deep Dive

Towards Human Motion World Models via Executable Behaviour Representations

Published on 2026-08-24

The rapid growth of human‑centric AI demands models that not only recognise actions but also understand the intentionality behind motion. The recent arXiv paper “Towards Human Motion World Models via Executable Behaviour Representations” introduces a novel solution: the ExAct domain‑specific language (DSL), which encodes human motions as underspecified programs that can be compiled into reward models for zero‑shot policy inference.

Key Contributions

• Executable Behaviour Representations: Motions are expressed as modular code blocks, enabling direct execution and evaluation.
• Zero‑Shot Policy Inference: By compiling ExAct programs to reward functions, new actions can be performed without additional training data.
• Compositionality: Individual policies can be combined seamlessly, supporting complex behaviour synthesis.
• Empirical Validation: Demonstrated improvements in data efficiency for human action segmentation and anomaly detection.

Why ExAct Matters for UBOS

At UBOS we specialise in building scalable, data‑efficient AI pipelines. ExAct aligns perfectly with our mission:

1. Modular Design: Engineers can reuse motion primitives across projects, reducing development time.
2. Zero‑Shot Capabilities: New behaviours can be deployed instantly, accelerating product iterations.
3. Interpretability: The program‑like representation offers clear insight into why a model predicts a certain motion.

Technical Overview

ExAct programs consist of three core elements:

1. Variables that capture latent intent (e.g., speed, direction).
2. Operators that define motion primitives such as walk(), reach(), or grasp().
3. Constraints that encode physical plausibility and task‑specific goals.

These programs are compiled into a reward model R(s,a) that can be queried by any reinforcement‑learning policy. The reward model is learned from a modest set of motion‑capture demonstrations, yet it generalises to unseen actions through the compositional nature of the DSL.

Performance Highlights

Getting Started with ExAct on UBOS

Developers can explore the DSL and try the reference implementation here:

• ExAct DSL Documentation
• UBOS SDK – Motion World Models

Conclusion

The ExAct DSL opens a new frontier for human motion modelling by marrying programmatic expressiveness with reinforcement‑learning execution. For organisations seeking to build robust, data‑efficient motion‑aware AI, ExAct provides a clear, extensible pathway.

Ready to prototype your own executable motion models? Visit UBOS and let our team help you integrate ExAct into your next project.

Tags: human motion modeling, executable behaviour, ExAct DSL, zero-shot policy inference, action segmentation, anomaly detection, UBOS

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.