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

arXiv 2026 61.1 method, system

TLDR

μ0 predicts 3D interaction traces from video, enabling scalable robot learning without action labels.

Reasoning

The paper introduces a novel world model that predicts 3D trajectories for interaction points, offering a compact and embodiment-agnostic alternative to pixel or action models. Strengths include scalability from diverse video sources and competitive downstream policy performance, but weaknesses include reliance on keypoint extraction and lack of explicit real-world validation in the abstract.

Read-first score

Read-first score 61.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

93.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.93310138

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,evaluation,experiment,result

Topical relevance 29%

65.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 430.

Keyword Scores

world model

interactive world model

world dynamics prediction

generative world model

world simulator

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

3D interaction-trace world model that predicts smooth 3D trajectories for salient interaction points instead of dense pixels or embodiment-specific actions
TraceExtract system for automatic extraction of 3D supervision from diverse video sources via keypoint selection, globally aligned traces, and hierarchical language captions
Modular trace expert using B-spline control points to represent and predict future traces
Embodiment-agnostic motion interface enabling cross-embodiment transfer without action labels
Action-free pretraining that achieves competitive performance with action-supervised VLA models like π_0

Methodology

μ_0 combines a pretrained vision-language backbone with a modular trace expert that represents each query via B-spline control points and predicts future 3D traces. Training data is generated by the TraceExtract system, which automatically selects keypoints, constructs globally aligned traces, and associates motion segments with hierarchical language captions from diverse video sources. Evaluation includes 2D and 3D trace prediction against baselines and downstream trace-conditioned policy performance on robot embodiments.

Key Results

μ_0 outperforms baselines in both 2D and 3D trace prediction, including trace prediction models and tokenized VLM methods. Trace-conditioned policies, despite action-free pretraining, achieve performance competitive with VLA models pretrained with action supervision, such as π_0.

Resources

Paper