Driving · 2026-10-30

Continuous-Action World Models: Learning Cross-Embodiment Spatiotemporal Dynamics Priors

Liangyu Li, Qingwen Liu

arXiv:2610.40401PDF

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

TL;DR

CoAWM (Continuous-Action World Models) learns a shared latent function space for robot actions across embodiments and execution frequencies. It uses two stages: (A) pretraining an action-function encoder/decoder with action reconstruction and consistency across temporal sampling views, and (B) aligning the learned function space to a world-model backbone conditioned on observations, language, and robot state, then decoding actions via time queries at desired execution frequencies.

Why it matters

The authors aim to make robot world models reusable across different robot bodies and control rates by separating (i) learning shared spatiotemporal dynamics priors in a continuous action-function space from (ii) selecting actions from the current scene/task context. Time queries let the same predicted function produce actions at different execution frequencies.

Method

  • Stage A: pretrain an action-function encoder/decoder from timestamped end-effector trajectories across embodiments and two sampling rates; learn dynamics priors via action reconstruction and consistency across temporal sampling views.
  • Stage B: connect the pretrained latent function space to a visual world-model backbone (VLM features, predicted future visual features, and proprioception + language) with a prediction head; align predicted latent functions and supervise via decoded-action errors.
  • Inference: predict one latent function per replanning step, anchor it to the current pose, and decode end-effector actions by querying the function at time points determined by the desired execution frequency.

Limitation

The authors report that “substantial jerk tails remain over longer execution windows,” and state they will “further address action stability during sustained execution and replanning in subsequent optimization.”

Abstract (from arXiv)

We introduce Continuous-Action World Models (CoAWM), a framework for learning cross-embodiment spatiotemporal dynamics priors that support robot action generation across embodiments and execution frequencies. At its core is a shared latent function space that captures the spatial and temporal structure of robot actions. We learn this space through two stages. First, function pretraining brings trajectories from different embodiments and sampling frequencies into a common representation, learning shared dynamics priors through action reconstruction and consistency across temporal sampling views. By accommodating heterogeneous robot trajectories, this stage broadens the range of data sources available for pretraining and facilitates scaling to larger and more diverse datasets. Second, we connect the pretrained function space to a world-model backbone, enabling observations, language instructions, and the current robot state to condition the prediction of a latent action function. A single decoder takes any query time and returns the corresponding end-effector action. By separating shared dynamics priors from task-conditioned action selection and temporal sampling, our approach provides a unified foundation for transferring robot experience across embodiments and control frequencies.

Related papers

Publication record

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

First public release: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC

preprint

Continuous-Action World Models: Learning Cross-Embodiment Spatiotemporal Dynamics Priors

Liangyu Li, Qingwen Liu

Preprint posted: 2026-10-29 · Original record ↗

arxiv
2610.40401
Explicit source versions (1)
  • v1 · 2026-10-29
  1. Preprint available

    Source date: 2026-10-29 · First detected: 2026-10-30 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

We introduce Continuous-Action World Models (CoAWM), a framework for learning cross-embodiment spatiotemporal dynamics priors that support robot action generation across embodiments and execution frequencies. At its core is a shared latent function space that captures the spatial and temporal structure of robot actions. We learn this space through two stages. First, function pretraining brings trajectories from different embodiments and sampling frequencies into a common representation, learning shared dynamics priors through action reconstruction and consistency across temporal sampling views. By accommodating heterogeneous robot trajectories, this stage broadens the range of data sources available for pretraining and facilitates scaling to larger and more diverse datasets. Second, we connect the pretrained function space to a world-model backbone, enabling observations, language instructions, and the current robot state to condition the prediction of a latent action function. A single decoder takes any query time and returns the corresponding end-effector action. By separating shared dynamics priors from task-conditioned action selection and temporal sampling, our approach provides a unified foundation for transferring robot experience across embodiments and control frequencies.