Multimodal Fusion and Vision–Language Models: A Survey for Robot Vision

Xiaofeng Han · Shunpeng Chen · Zenghuang Fu · Zhe Feng · Lue Fan · Dong An · Zhangwei Wang · Li Guo · Weiliang Meng* · Xiaopeng Zhang · Rongtao Xu* · Shibiao Xu

This repository tracks research on multimodal fusion and vision–language models (VLMs) for robot vision, covering semantic scene understanding, 3D perception, SLAM, navigation & localization, and manipulation. We also summarize datasets, metrics, challenges (e.g., cross-modal alignment, efficient fusion, real-time deployment), and future directions.

⭐ Share us a ⭐

Share us a ⭐ if you're interested in this repo. We will continue to track relevant progress and update this repository.

News

[2025-09-03] Camera-ready accepted by Information Fusion (Vol. 126, 2026). DOI: 10.1016/j.inffus.2025.103652.

[2025-09-01] Featured by Embodied Intelligence Hub — article recap of our survey and repo. Read the post »

📖 Introduction

The overview figure illustrates the overall framework of multimodal fusion and VLMs for robot vision:

Table of Contents

• MF-RV"
  • News
  • Overview
  • Table of Contents
  • Related Surveys
  • Awesome Benchmarks
  • Embodied Large Language Models

Related Surveys

📊Awesome Benchmarks

Scene Understanding Datasets

Datasets	Scene	Multimodal Data	Venue	Year
360+x	Indoor/Outdoor	Video/Audio	CVPR	2024
ScanQA	Indoor	RGB/Text	CVPR	2022
Hypersim	Indoor	RGB/Depth	ICCV	2021
NuScenes	Urban street	RGB/Lidar/Radar	CVPR	2020
Waymo	Outdoor	RGB/Lidar	CVPR	2020
Semantickitti	Urban street	RGB/Lidar	ICCV	2019
Matterport3D	Indoor	RGB/Depth	arxiv	2017
ScanNet	Indoor	RGB/Depth	CVPR	2017
Cityscapes	Urban street	RGB/Depth	CVPR	2016
NYUDv2	Indoor	RGB/Depth	ECCV	2012

Robot Manipulation Datasets

Datasets	Core Modalities	Data Scale	Main Application
DROID	RGB, Depth, Text	76,000 trajectories	Multi-task scene adaptation
R2SGrasp	RGB-D, Point Cloud	64,000 RGB-D images	Grasp detection
RT-1	RGB, Text	130,000 trajectories	Real-time task control
Touch and Go	RGB, Tactile	3,971 virtual object models, 13,900 tactile interactions	Cross-modal perception
VisGel	GelSight Tactile, RGB	12,000 tactile interactions	Tactile-enhanced manipulation
ObjectFolder 2.0	RGB, Audio, Tactile	1,000 virtual object models	Virtual-to-reality transfer
Grasp-Anything-6D	Point Cloud, Text	1M point cloud scenes	Language-driven grasping
Grasp-Anything++	Point Cloud, Text	1M samples, 10M instructions	Fine-grained manipulation
Open X-Embodiment	RGB, Depth, Text, Multi-robot Data	Aggregated data from multiple institutions	Cross-robot system generalization

• Grasp-Anything++: Language-driven Grasp Detection(CVPR, 2024) [paper]
• Grasp-Anything-6D: Language-Driven 6-DoF Grasp Detection Using Negative Prompt Guidance(ECCV, 2024) [paper]
• Real-to-Sim Grasp: Rethinking the Gap between Simulation and Real World in Grasp Detection(CORL, 2024) [paper]
• Open X-Embodiment: Robotic Learning Datasets and RT-X Models(ICRA, 2024) [paper]
• DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset(ICRA, 2024) [paper]
• Objectfolder 2.0: A multisensory object dataset for sim2real transfer (CVPR, 2022) [paper]
• Touch and Go: Learning from Human-Collected Vision and Touch Supplementary Material (NuerIPS, 2022) [paper]
• Connecting Touch and Vision via Cross-Modal Prediction (CVPR, 2019) [paper]
• Rt-1: Robotics transformer for real-world control at scale (arXiv, 2022) [paper]

Embodied Navigation Datasets

Dataset	Modalities	Unique Feature
Matterport3D	RGB-D, Semantic Annotations	Foundational dataset for navigation
R2R	RGB-D, Natural Language	Vision-and-Language Navigation
REVERIE	RGB-D, Object Annotations	Combines object grounding tasks
CVDN	RGB-D, Dialog	Introduces multi-turn interactions
SOON	RGB-D, Natural Language	Coarse-to-fine target localization
R3ED	Point Cloud, Object Labels	Real-world sensor-based data

Manipulation Benchmarks

Title	Venue	Date
Manipulation in Home Environment
HomeRobot: Open-Vocabulary Mobile Manipulation	CoRL 2023	2023-06-20
ALFRED: A Benchmark for Interpreting Grounded Instructions for Everyday Tasks	CVPR 2020	2019-12-03
Manipulation in On-Table Environment
OBSBench: Point Cloud Matters: Rethinking the Impact of Different Observation Spaces on Robot Learning	NeurIPS 2024	2024-02-04
LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning	NeurIPS 2023	2023-06-05

Benchmark	Simulator	# Tasks	Real-World Reproducibility	Applicable Algorithms	Key Evaluation Metrics
RLBench	RLBench	100+	✗	RL, IL, Traditional Control	Task Success Rate, Trajectory Efficiency, Task Completion Time
GemBench	RLBench	44	✗	RL, IL, VLM-based	Zero-shot Task Success, Object Recognition, Generalization
VLMbench	RLBench	8	✗	RL, VLM-based	Task Execution Success, Compositional Generalization
KitchenShift	Isaac Sim	7	✗	IL, RL	Performance Under Domain Shifts, Task Success Rate
CALVIN	PyBullet	34	✗	RL, IL	Long-Horizon Task Success, Multi-Task Adaptability
COLOSSEUM	RLBench	20	✓	RL, IL	Robustness to Perturbations, Multi-Task Learning Performance
VIMA	Ravens	17	✗	RL, VLM-based	Zero-shot Success, Multi-Modal Task Performance

Embodied Large Language Models

• Rt-2: Vision-language-action models transfer web knowledge to robotic control (arXiv, 2023) [paper]
• Open x-embodiment: Robotic learning datasets and rt-x models (arXiv, 2023) [paper]
• Rt-h: Action hierarchies using language (arXiv, 2024) [paper]
• Autort: Embodied foundation models for large scale orchestration of robotic agents (arXiv, 2024) [paper]
• Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation (arXiv, 2024) [paper]
• Voxposer: Composable 3d value maps for robotic manipulation with language models (arXiv, 2023) [paper]
• Rdt-1b: a diffusion foundation model for bimanual manipulation (arXiv, 2024) [paper]
• Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation (arXiv, 2024) [paper]
• Copa: General robotic manipulation through spatial constraints of parts with foundation models (arXiv, 2024) [paper]
• OpenVLA: An Open-Source Vision-Language-Action Model (arXiv, 2024) [paper]
• Palm-e: An embodied multimodal language model (arXiv, 2023) [paper]
• Π0: A Vision-Language-Action Flow Model for General Robot Control (arXiv, 2024) [paper]

📖 Citation

If you find our survey and repository useful for your research, please consider citing our paper:

@article{han2025multimodal,
  title={Multimodal fusion and vision-language models: A survey for robot vision},
  author={Han, Xiaofeng and Chen, Shunpeng and Fu, Zenghuang and Feng, Zhe and Fan, Lue and An, Dong and Wang, Changwei and Guo, Li and Meng, Weiliang and Zhang, Xiaopeng and others},
  journal={Information Fusion},
  pages={103652},
  year={2025},
  publisher={Elsevier}
}