$μ_0$: A Scalable 3D Interaction-Trace World Model

2026-06-11

AI Summary

This work addresses the limited cross-platform generalization of existing world models, which often rely on pixel reconstruction or embodied action labels. The authors propose μ₀, a novel world model that introduces 3D interaction trajectories as a transferable intermediate representation. By predicting smooth 3D motion trajectories of keypoints associated with objects, tools, and hands, μ₀ establishes a compact, embodiment-agnostic action interface. The architecture integrates a vision-language backbone with a B-spline trajectory expert module and leverages the TraceExtract system to automatically extract aligned 3D trajectories and hierarchical language annotations from multi-source videos for pretraining. Experiments demonstrate that μ₀ outperforms prior methods in both 2D and 3D trajectory prediction. Remarkably, when frozen and combined with downstream action experts, μ₀ achieves manipulation performance on par with action-supervised vision-language-action (VLA) models without requiring any action-specific pretraining.

Abstract

World models that capture how actions induce physical change enable scalable robot learning without reliance on embodiment-specific action labels. Pixel-space video models provide broad visual priors but expend model capacity on dense appearance reconstruction, while direct action models require embodiment-specific labels that hinder scalability. We present $μ_0$, a scalable world model based on 3D traces. Rather than predicting dense pixels or directly modeling actions, $μ_0$ forecasts smooth 3D trajectories for salient interaction points such as objects, tools, hands, and contact regions, yielding a compact, embodiment-agnostic motion interface. To enable training from diverse video sources, our TraceExtract system automatically extracts 3D supervision by selecting keypoints, constructing globally aligned traces, and associating motion segments with hierarchical language captions. This TraceExtract supervision pretrains $μ_0$ by combining a pretrained vision-language backbone with a modular trace expert, which represents each query via B-spline control points and predicts future traces. Experiments show that $μ_0$ outperforms baselines in both 2D and 3D trace prediction, including trace prediction models and tokenized VLM methods. Because $μ_0$ is frozen and reusable, it can be paired with action experts for downstream robot embodiments. Despite action-free pretraining, the resulting trace-conditioned policies achieve performance competitive with VLA models pretrained with action supervision, such as $π_0$. These results establish 3D traces as a scalable and transferable representation for cross-embodiment manipulation.

Problem

• world model
• 3D interaction traces
• embodiment-agnostic
• scalable robot learning
• action-free pretraining

Innovation

• 3D interaction traces
• embodiment-agnostic world model
• action-free pretraining
• trace prediction
• scalable robot learning