Latent Dynamics · 2026-10-19
Unifying Embodied Foundation Models via a Model-to-Policy Formulation
Xinyi Hong, Mingyang Ren
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 propose an architecture-agnostic probabilistic formulation for embodied foundation models using stochastic kernels, explicitly connecting model output distributions to action realization kernels and thus to closed-loop, task-conditioned robot policies. They show how pre-training and post-training objectives relate to deployment, and instantiate the framework for OpenVLA (autoregressive action-token prediction) and DreamZero (joint video-action generation).
Why it matters
The work provides a common probabilistic language linking heterogeneous embodied model architectures to the policies they induce in closed-loop interaction. By separating the model output distribution fθ from the action-realization kernel qθ, the formulation clarifies how changes during pre-training and post-training propagate to deployed robot behavior.
Method
- Formulate an embodied foundation model as an output distribution fθ mapping interaction history and task condition to probabilities over an output space Y.
- Define a measurable action-realization kernel qθ that maps model outputs to executable robot actions, inducing a task-conditioned policy πθ in closed loop.
- Connect offline/online pre-training and post-training losses to deployment by combining fθpost with qθpost to sample outputs y_t and realize actions a_t.
Abstract (from arXiv)
Vision-language-action (VLA) models and world-action models (WAMs) have emerged as promising approaches to robot learning and control. These models integrate multimodal perception and action generation, but differ in their architectures, output representations, and learning objectives. In this paper, we present a unified, architecture-agnostic probabilistic formulation for embodied foundation models based on stochastic kernels. Our formulation has two key components: (i) we characterize embodied foundation models through conditional output distributions and action-realization kernels, which together induce task-conditioned robot policies; and (ii) we connect pretraining and post-training to policy deployment, accommodating model-specific pretraining objectives and offline or online post-training strategies. We instantiate the formulation for OpenVLA and DreamZero, covering autoregressive action-token prediction and joint video-action generation.
Related papers
- Forward Dynamics Models
- PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
- UniWALM: Unified World Action Language Model for Embodied Robot Control
- What Should Robot Policies Learn to Predict? Rethinking Supervision Targets in Vision-Language-Action Learning
- La-JEPA: Compact Robot Policies with Latent World Modeling
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-16 · Indexed by Pub-AI: 2026-10-19 UTC
Unifying Embodied Foundation Models via a Model-to-Policy Formulation
Preprint posted: 2026-10-16 · Original record ↗
- arxiv
- 2610.23053
Explicit source versions (1)
- v1 · 2026-10-16
- Preprint available
Source date: 2026-10-16 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Vision-language-action (VLA) models and world-action models (WAMs) have emerged as promising approaches to robot learning and control. These models integrate multimodal perception and action generation, but differ in their architectures, output representations, and learning objectives. In this paper, we present a unified, architecture-agnostic probabilistic formulation for embodied foundation models based on stochastic kernels. Our formulation has two key components: (i) we characterize embodied foundation models through conditional output distributions and action-realization kernels, which together induce task-conditioned robot policies; and (ii) we connect pretraining and post-training to policy deployment, accommodating model-specific pretraining objectives and offline or online post-training strategies. We instantiate the formulation for OpenVLA and DreamZero, covering autoregressive action-token prediction and joint video-action generation.
