Koopman Dreamer Enhances Stable World-Model Imagination for Control

Jiaqi Li, Xinglong Zhang, Haibin Xie, Yixing Lan, Wei Pan, Xin Xu· July 23, 2026 View original

Summary

Koopman Dreamer is a new world model that improves stability in long-horizon latent rollouts for continuous control tasks. It uses spectrally constrained latent dynamics with rotation-scaling blocks to better manage error accumulation and modal persistence.

This research introduces Koopman Dreamer, an advancement in world models designed for continuous control tasks. It addresses the challenge of error accumulation and limited control over modal persistence in long-term simulations by incorporating a spectrally constrained deterministic latent dynamics core. The model utilizes Koopman-inspired two-dimensional rotation-scaling blocks with bounded radii, enabling it to represent damping, rotation, and near-periodic modes more effectively. The system also integrates linear and low-rank bilinear action terms for global and state-dependent control, complemented by stochastic-state modulation for local corrections. To bridge the gap between training and imagination, Koopman Dreamer employs a combination of posterior-conditioned EMA teacher targets and various consistency objectives. Experimental results demonstrate improved stability in long-horizon latent rollouts and superior closed-loop control performance on tasks requiring high-quality multi-step imagination, such as DeepMind Control Suite and UAV-LiDAR navigation.

Why it matters

Professionals in robotics, autonomous systems, and AI development can leverage this research to build more stable and reliable AI agents capable of planning over longer time horizons.

How to implement this in your domain

  1. 1Explore integrating spectrally constrained latent dynamics into existing model-based reinforcement learning architectures.
  2. 2Evaluate the Koopman Dreamer approach for tasks requiring long-term planning and stable predictions in simulated environments.
  3. 3Adapt the two-dimensional rotation-scaling blocks to represent specific system dynamics in your control applications.
  4. 4Implement the proposed multi-step rollout and observation-prediction objectives to improve model consistency.

Who benefits

RoboticsAutonomous VehiclesAerospaceIndustrial Automation

Key takeaways

  • Koopman Dreamer improves the stability of long-horizon latent rollouts in world models.
  • It uses spectrally constrained dynamics to better manage error accumulation and modal persistence.
  • The model shows stronger closed-loop control performance on tasks relying on multi-step imagination.
  • This approach offers a path to more reliable and robust AI agents for continuous control.

Original post by Jiaqi Li, Xinglong Zhang, Haibin Xie, Yixing Lan, Wei Pan, Xin Xu

"arXiv:2607.19719v1 Announce Type: new Abstract: Latent world models improve sample efficiency in continuous control by optimizing policies over imagined latent trajectories, but common neural transitions offer limited direct control over modal persistence and error accumulation i…"

View on X

Originally posted by Jiaqi Li, Xinglong Zhang, Haibin Xie, Yixing Lan, Wei Pan, Xin Xu on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses