Latent Dynamics · 2026-11-10

Dynamics-Informed Action Learning for Goal-Conditioned Latent Planning

Yinuo Jiang, Guangzhi Wang, Yiran Wang, Junfu Pu, Cheng Cheng, Honghao Chen, Cheng Cheng, Ying Shan

arXiv:2611.10660PDF

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

TL;DR

DIAL (Dynamics-Informed Action Learning) trains latent dynamics with rollout alignment and trains a lightweight inverse dynamics head with multi-step inverse supervision. At test time, the inverse and forward dynamics are coupled to generate a goal-directed action sequence that initializes only the mean of CEM’s proposal distribution; standard CEM sampling/refinement then improves the plan. The authors report DIAL outperforms prior latent world-model and planning baselines on four visual control benchmarks.

Why it matters

The paper targets a gap between how latent dynamics are trained (teacher-forced one-step prediction) and how they’re used for goal-conditioned latent planning (autoregressive rollouts with accumulated drift), and also a gap in how sampling-based planners initialize action search (goal-agnostic proposals). DIAL directly connects action-conditioned latent dynamics learning to goal-directed CEM initialization via an inverse dynamics head.

Method

  • Rollout alignment: train the forward latent dynamics model on autoregressively generated latent contexts (depth L), reducing mismatch between teacher forcing and planning-time rollouts.
  • Multi-step inverse supervision: train a lightweight inverse dynamics head to recover the initiating action from latent-state transitions; apply a frozen inverse head to model-generated transitions to encourage preservation of action-relevant information.
  • Dynamics-informed planning: autoregressively couple inverse and forward dynamics to produce a goal-directed action sequence, used to initialize only the mean of CEM’s Gaussian proposal; then run standard CEM sampling/refitting for refinement.

Limitation

The inverse head is not a uniformly accurate action predictor at larger temporal gaps: the authors report degraded inverse prediction and note its predictions provide only a coarse goal-directed initialization that is subsequently refined by CEM.

Abstract (from arXiv)

Latent world models enable visual control by planning action sequences in compact representation spaces, but a gap remains between how latent dynamics are learned and how they are used for planning. First, teacher-forced training does not expose the dynamics model to the recursively predicted latent states encountered during planning. Second, sampling-based planners typically initialize action search from goal-agnostic proposals, without exploiting the action-conditioned dynamics learned by the world model. We introduce Dynamics-Informed Action Learning (DIAL) to bridge these two gaps. During training, rollout alignment exposes the forward model to its own autoregressive latent predictions, while multi-step inverse supervision trains a lightweight inverse dynamics head to infer actions from latent-state transitions. We further apply the inverse head to model-generated transitions with its parameters held fixed, encouraging predicted rollouts to retain action-relevant information. At test time, DIAL autoregressively couples the inverse and forward dynamics to construct a dynamics-informed action sequence toward the goal, which initializes the mean of the Cross-Entropy Method (CEM) proposal distribution. The subsequent CEM search then refines this initialization using the standard sampling and updating procedure. Experiments across four visual control benchmarks demonstrate that DIAL outperforms existing latent world models and planning baselines in both standard and long-horizon scenarios.

Related papers

Publication record

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

First announced by arXiv: 2026-11-10 · Indexed by Pub-AI: 2026-11-10 UTC

preprint

Dynamics-Informed Action Learning for Goal-Conditioned Latent Planning

Yinuo Jiang, Guangzhi Wang, Yiran Wang, Junfu Pu, Cheng Cheng, Honghao Chen, Cheng Cheng, Ying Shan

First announced by arXiv: 2026-11-10 · Original record ↗

arxiv
2611.10660
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-10 · First detected: 2026-11-10 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.10660v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

Latent world models enable visual control by planning action sequences in compact representation spaces, but a gap remains between how latent dynamics are learned and how they are used for planning. First, teacher-forced training does not expose the dynamics model to the recursively predicted latent states encountered during planning. Second, sampling-based planners typically initialize action search from goal-agnostic proposals, without exploiting the action-conditioned dynamics learned by the world model. We introduce Dynamics-Informed Action Learning (DIAL) to bridge these two gaps. During training, rollout alignment exposes the forward model to its own autoregressive latent predictions, while multi-step inverse supervision trains a lightweight inverse dynamics head to infer actions from latent-state transitions. We further apply the inverse head to model-generated transitions with its parameters held fixed, encouraging predicted rollouts to retain action-relevant information. At test time, DIAL autoregressively couples the inverse and forward dynamics to construct a dynamics-informed action sequence toward the goal, which initializes the mean of the Cross-Entropy Method (CEM) proposal distribution. The subsequent CEM search then refines this initialization using the standard sampling and updating procedure. Experiments across four visual control benchmarks demonstrate that DIAL outperforms existing latent world models and planning baselines in both standard and long-horizon scenarios.