PitchBook-新兴空间简报_机器人基础模型（英）-2025_8页_458kb

Summary of Robotic Foundation Models from PitchBook

Definition and Background

Robotic Foundation Models (RFMs) are AI models acting as unified "brains" for robots, integrating vision, language, and motion capabilities. They are pretrained on vast datasets of internet-scale text, imagery, and robotic experiences, enabling generalized task performance with human-like adaptability. RFMs evolved from earlier deterministic robotics systems and were inspired by advances in AI foundation models, with early projects like DeepMind's Gato demonstrating their potential in the 2020s.

Key Technologies

• Multimodal transformer architectures support unified encoding and decoding of inputs (e.g., Google's RT-2 translates visuals and language to control signals).
• Data collection involves fleet learning (e.g., Ambi Robotics using 20 million images), simulation-to-reality pipelines (e.g., NVIDIA Omniverse), and self-supervised learning techniques.
• Hardware-software codesign relies on powerful GPUs/TPUs for training and focuses on model compression and edge hardware for real-time deployment on robots.

Applications

RFMs are applied in various industries:

• Industrial automation and manufacturing enable flexible robotics and human collaboration.
• Warehouse logistics improve handling of diverse items through systems like Covariant's Brain.
• Domestic assistance allows robots to perform tasks such as tidying or cooking.
• Defense sector enhances autonomous systems for military robots and reconnaissance.

Limitations

Challenges include scarcity of high-quality training data, poor generalization to novel environments, high computational latency, safety risks from unpredictable behavior, and operational issues like integration with legacy systems and talent shortages.

Recent Deal Activity and Market Outlook

There has been significant investment, with over $2.2 billion deployed in RFM startups in 2025, a 7x increase from 2023. Notable investments include Physical Intelligence's $400 million round and Scout AI's defense-focused seed funding. Strategic acquisitions, such as Intrinsic acquiring Vicarious, reflect tech giants' interest. The market is expected to grow, with a projected total addressable market of hundreds of billions, expanding into warehouses, manufacturing, consumer robotics, and defense, driven by potential cost reductions and new business models like robot-as-a-service.

Quantitative Overview

Leading companies by funding include Physical Intelligence ($470M total raised), Covariant ($245M), and Figure AI ($2,345M). Exit opportunities for Physical Intelligence are high, while companies like Physical Intelligence and Chef Robotics lead in active patents. Venture capital activity shows strong interest from investors like BOLD Capital Partners and Amazon.