California robotics startup FieldAI has established its unicorn status after raising $405 million in a two-stage funding round this August, pushing its valuation to roughly $2 billion. The haul signals strong investor confidence in the company’s push to bring “embodied AI,” artificial intelligence designed for physical systems, into mainstream deployment.
Formerly known as AI for Humanity, FieldAI has staked its future on what it calls “Field Foundation Models (FFMs).” Unlike AI software retrofitted for robotics, FFMs are built to be “embodiment-agnostic,” using machine learning architectures that account for physics, risk, and uncertainty from the ground up.
The company says this design allows robots, from humanoids and quadrupeds to drones, rovers, and industrial vehicles, to operate in messy, real-world environments without the use of GPS, pre-mapped data, or human intervention. In short, FieldAI wants to make AI systems that can think and move in the field as reliably as they do in the lab.
FieldAI attracting notable investor names
FieldAI has raised a total of $405 million across two funding rounds, starting with $91 million in late 2024 and followed by a $314 million round announced on August 20, 2025. The latest raise pushed the company’s valuation close to $2 billion and was co-led by Bezos Expeditions, Prysm Capital, and Temasek, signaling strong support from both established venture firms and strategic investors.
The startup’s supporters now include Khosla Ventures, Canaan Partners, Intel Capital, NVentures (NVIDIA’s venture arm), BHP Ventures, and Emerson Collective, with earlier investments from Gates Frontier and Samsung. Collectively, this roster highlights broad confidence in FieldAI’s efforts to expand the reach of its FFM-enabled robots across multiple industries and applications.
From NASA research to a $2B robotics startup
FieldAI’s origins date back to 2016, when its founding team began developing autonomous systems through NASA’s BRAILLE project. The work focused on using robots to map planetary-like lava tubes without satellite data or preloaded maps. Between 2018 and 2022, the group advanced that research through DARPA’s Subterranean Challenge and RACER initiative, where their technology demonstrated adaptability and resilience in real-world field deployments.
Building on these results, the team formally launched FieldAI in 2023 with the goal of creating a general-purpose, deployable robotic intelligence, a “universal brain” for machines that can adapt across platforms and environments. This mission has attracted significant funding and positioned the company as a leading player in the development of embodied AI.
Today, FieldAI is valued at $2 billion and is scaling operations across the United States, Japan, and Europe. Its workforce includes veterans of NASA JPL, DARPA, DeepMind, Google Brain, NVIDIA, Qualcomm, Tesla, Zoox, Cruise, Toyota Research Institute, SpaceX, Boston Dynamics, Amazon, Microsoft, and other major research labs and companies. The company now provides solutions for diverse applications, including construction, urban operations, industrial and energy, federal projects, and many more.
FFM: FieldAI’s approach to embodied AI
FieldAI describes its “Field Foundation Models” (FFM) as a new class of foundation models built specifically for robots in complex, real-world environments. Unlike traditional systems that rely on GPS, pre-mapped data, or rigid programming, FFM integrates multimodal inputs, such as vision, LiDAR, and text, with large-scale deployment datasets to help robots adapt to unfamiliar situations.
The company says its models embrace uncertainty rather than avoid it, allowing robots to anticipate outcomes, make faster and safer decisions, and adjust in unstructured or hazardous conditions. A single deployable “universal brain,” as FieldAI describes it, is designed to run entirely on the edge, reducing dependence on heavy cloud infrastructure and enabling rapid deployment across industries.
Foundational models in robotics
The robotics field has increasingly looked to “foundation models,” which are large-scale, pre-trained AI systems, as a path toward general-purpose capabilities. A 2023 survey co-authored by FieldAI researchers Ali Agha, Shayegan Omidshafiei, and Dong-Ki Kim, titled “Toward General-Purpose Robots via Foundation Models,” categorized current efforts into two approaches: adapting existing vision and language models for robotic control, and building robotics-specific foundation models (RFMs) trained directly on robot data. FieldAI’s FFMs fall into the latter category.
These models aim to address key hurdles in embodied AI, such as generalization, data scarcity, and the need for uncertainty-aware decision-making. Foundation models also offer improved “zero-shot” adaptability across diverse tasks and environments, useful integration of limited real-world datasets, and potentially better task specification through language or other modalities. Notably, uncertainty and safety remain open challenges, areas FieldAI is placing particular emphasis on.
Toward AGI in robotics: the next frontier in AI
The rise of foundation models in robotics is fueling discussions over artificial general intelligence (AGI). Fei-Fei Li, often called the “godmother of AI” and a Stanford professor who co-founded World Labs, has argued that progress toward AGI requires spatial intelligence, the ability to build predictive world models grounded in 3D structure and physics. Li called it the next frontier in AI. Language-only systems, she noted, lack grounding in the physical world and therefore fall short of this goal.
The embodied AI landscape
The push for embodied AI has drawn startups, tech giants, and research labs into a rapidly intensifying field. New entrants such as Skild AI, Physical Intelligence, Genesis AI, Collaborative Robotics (Cobot), and RLWRLD are all building universal or robotics-specific models, often backed by significant venture capital and partnerships. Meanwhile, Google DeepMind and NVIDIA are pursuing their respective projects ranging from Gemini Robotics’ vision-language-action systems to Genie’s 3D world models with embodied agents and Isaac GR00T’s dual-system humanoid framework. Each effort reflects a growing consensus that foundation models represent a transformative shift in robotics, much like how large language models have reshaped natural language AI.
FieldAI stands out for its explicit focus on risk- and uncertainty-aware foundation models. Described as the first of its kind among robotic foundation models, its FFMs emphasize resilience and adaptability in unstructured, real-world conditions and are already being actively deployed across customer sites worldwide. By emphasizing risk-aware autonomy and scalable deployment, the company is carving out a distinct role in the fast-growing embodied AI market.
Uncertainty-aware navigation in real-world environments
Autonomous robots can face significant challenges when operating in rough or cluttered terrain. Traversability estimation is often hindered by sensor noise, occlusions, and dynamic environmental changes, as well as the difficulty of collecting sufficient real-world training data. Conventional approaches that depend on high-precision scanners, simulations, or odometry are limited in their ability to capture uncertainties and environmental complexity, resulting in unreliable navigation and restricted deployment in real-world scenarios.
U.S. Pat. App. Pub. No. 2025/0252306 describes a machine learning–based framework designed to overcome these limitations. The invention generates “optimum-fidelity” scan data and synthetic point clouds to create high-quality training datasets, simulates multiple viewpoints and introduces noise to replicate the performance of low-resolution sensors. Using these inputs, the system predicts terrain features such as slope, roughness, and step height, expressing traversability as probability distributions rather than fixed outcomes. This uncertainty-aware approach enables robots to adapt their decision-making in real time, improving safety and reliability in complex environments.
The application builds directly on the 2024 UNRealNet paper, in which the inventors demonstrated a label-free, uncertainty-aware approach to traversability estimation. Their work showed how synthetic point clouds derived from high-fidelity scans could overcome occlusions, odometry drift, and sparse sensing, producing robust, robot-agnostic traversability estimates, aspects which were later formalized in the patent filing.
The patent application, titled “System and method for uncertainty-aware traversability estimation with optimum-fidelity scan data,” was filed on February 5, 2025, following a provisional filing in 2024, and published on August 7, 2025. The application is currently pending. The patent application lists Samuel Triest, David Fan, and Ali Agha as inventors. Legal representation was provided by Stephen Becker, Gregory Montone, Bret Petersen et al. from NovoTechIP International PLLC. Although not named in the filing, Sebastian Scherer contributed to the related research as a co-author of UNRealNet alongside other FieldAI team members.
FieldAI’s potential IP strategy and paths forward
FieldAI’s first patent filing highlights its strategy of embedding risk-awareness, adaptability to diverse sensors, and generalizable learning into its Field Foundation Models (FFMs). These innovations align with the company’s vision of building scalable, general-purpose embodied AI capable of operating reliably in the physical world. Securing protection for this work marks an important step in strengthening FieldAI’s position in the fast-emerging physical AI market.
Looking ahead, the company’s IP strategy could evolve in several ways. It may expand into a full patent portfolio covering FFM-related innovations, adopt an open-source stance to encourage ecosystem adoption, or pursue a hybrid approach that protects high-value architecture while sharing general-purpose components. Such positioning could prove valuable as the global embodied AI market, valued at $4.44 billion with a 39% CAGR through 2030, and the global robotics market, projected at $50.8 billion with a 9.49% CAGR through 2029, continue to grow. For now, FieldAI seems to be focused on its FFM research and development and on scaling its real-world deployment and operations, but the ‘306 application signals a clear recognition of IP as a driver of long-term differentiation.
Note: Thumbnail image is for illustrative purposes only and does not depict the actual robots of Field AI
