⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

3D-VLA: A 3D Vision-Language-Action Generative World Model

ICML 24 2024 61.3 method

TLDR

3D-VLA integrates 3D perception, reasoning, and action via a generative world model using LLM and diffusion models, improving embodied planning.

Reasoning

The paper introduces a novel 3D embodied foundation model that explicitly models world dynamics and generation, which is a strength. However, it only evaluates on held-in datasets without real-world validation, limiting evidence of practical applicability.

Read-first score

Read-first score 61.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Methodology quality 25%
80

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

Recency 8%
75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Topical relevance 42%
61.4

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 25%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

Methodology anchor

Rank sensitivity

Stability: volatile; rank range: 71.

Keyword Scores

world model
10
generative world model
10
world dynamics prediction
8
interactive world model
6
model-based reinforcement learning world model
4
world simulator
3
video world model
2

Deep Analysis

Innovations

  • Introduces 3D-VLA, a family of embodied foundation models that link 3D perception, reasoning, and action through a generative world model.
  • Uses interaction tokens to engage with the embodied environment.
  • Trains embodied diffusion models aligned into a 3D-based LLM for predicting goal images and point clouds.
  • Curates a large-scale 3D embodied instruction dataset from existing robotics datasets.

Methodology

3D-VLA is built on top of a 3D-based large language model (LLM) with a set of interaction tokens to engage with the embodied environment. It trains a series of embodied diffusion models and aligns them into the LLM for predicting goal images and point clouds. The model is trained on a curated large-scale 3D embodied instruction dataset extracted from existing robotics datasets.

Key Results

On held-in datasets, 3D-VLA significantly improves reasoning, multimodal generation, and planning capabilities in embodied environments, demonstrating potential for real-world applications.

Limitations

  • Evaluation is conducted only on held-in datasets, leaving generalization to out-of-distribution or real-world scenarios unaddressed.

Resources

Paper

Tags