Latent Dynamics · 2026-11-02

Learning Stabilizing Latent Feedback Controllers via Multi-Step Backward Losses

Paul Lutkus, Lars Lindemann, Stephen Tu

arXiv:2610.40931PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors study when learned latent dynamics models and encoders can support feedback controllers that provably stabilize the original system. For stabilizable LTI systems, they show that learning linear autoencoders and latent dynamics with sufficiently long multi-step “backward losses” (prediction-reconstruction) recovers structure needed for stabilization via output-feedback, provided the latent dimension captures all non-stable modes.

Why it matters

This work links world-model learning objectives to closed-loop stability guarantees. It shows that controller-relevant latent structure is not ensured by low forward prediction error alone: forward losses can admit minimizers that discard unstable modes, and single-step reconstruction can fail to yield latent models whose stabilizing controllers stabilize the original system. The paper provides an end-to-end latent stabilization guarantee for multi-step backward losses.

Method

  • Learn an encoder E and latent dynamics f_z (plus decoder D for backward losses) from trajectory data using multi-step prediction-reconstruction losses L_bwd; backward losses encourage latent trajectories to reconstruct full state trajectories over a horizon T>1.
  • Use the learned (E,f_z) to design an output-feedback controller by composing a latent controller K(z) with the encoder’s output map (linearization) for stabilization analysis.
  • For LTI systems, prove that sufficiently long horizon backward loss global minimizers satisfy approximate structural conditions (detectability-like and encoder invariance) that enable robust controller synthesis; show insufficiency of forward losses and single-step backward losses.
Abstract (from arXiv)

Can learned latent dynamics models support the design of feedback controllers that provably stabilize the original system? For linear time-invariant systems, we show that learning linear autoencoders and latent dynamics with multi-step prediction-reconstruction losses $\unicode{x2014}$ which we term "backward losses" as they encourage latent trajectories to reconstruct full state trajectories $\unicode{x2014}$ provably recovers the structure needed for stabilization. Specifically, a sufficiently long reconstruction horizon ensures that a minimum ${H}_\infty$-norm controller designed using the learned latent model stabilizes the original system through output-feedback via the corresponding encoder, provided the latent dimension is large enough to capture all non-stable modes of the underlying system. We also show why both reconstruction and a multi-step horizon are essential: forward prediction losses that require only low prediction error along latent trajectories admit global loss minimizers that discard unstable modes regardless of horizon length, while single-step reconstruction losses can yield latent models whose stabilizing controllers fail to stabilize the original system. Finally, experiments on a non-linear flexible-arm system illustrate the practical implications of our results for learning non-linear autoencoders and latent dynamics.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: Date not supplied · Indexed by Pub-AI: 2026-11-02 UTC

preprint

Learning Stabilizing Latent Feedback Controllers via Multi-Step Backward Losses

Paul Lutkus, Lars Lindemann, Stephen Tu

Source date: Date not supplied · Original record ↗

arxiv
2610.40931

    Imported from the existing registry. No new publication milestone asserted.

    Abstract · arxiv

    Can learned latent dynamics models support the design of feedback controllers that provably stabilize the original system? For linear time-invariant systems, we show that learning linear autoencoders and latent dynamics with multi-step prediction-reconstruction losses $\unicode{x2014}$ which we term "backward losses" as they encourage latent trajectories to reconstruct full state trajectories $\unicode{x2014}$ provably recovers the structure needed for stabilization. Specifically, a sufficiently long reconstruction horizon ensures that a minimum ${H}_\infty$-norm controller designed using the learned latent model stabilizes the original system through output-feedback via the corresponding encoder, provided the latent dimension is large enough to capture all non-stable modes of the underlying system. We also show why both reconstruction and a multi-step horizon are essential: forward prediction losses that require only low prediction error along latent trajectories admit global loss minimizers that discard unstable modes regardless of horizon length, while single-step reconstruction losses can yield latent models whose stabilizing controllers fail to stabilize the original system. Finally, experiments on a non-linear flexible-arm system illustrate the practical implications of our results for learning non-linear autoencoders and latent dynamics.