CtrlK
Field AI (also written FieldAI) is an American robotics software company based in Irvine, California, that develops foundation models specifically designed for autonomous robots operating in outdoor and unstructured industrial environments. The company was founded in 2023 by Ali-akbar Agha-Mohammadi and a team of researchers who previously led NASA's Jet Propulsion Laboratory (JPL) autonomy programs, including the DARPA Subterranean Challenge and DARPA RACER. Field AI's core product is a class of AI systems called Field Foundation Models (FFMs), which give robots a "physics-first" brain capable of navigating GPS-denied, unmapped terrain without human intervention. By August 2025, the company had raised $405 million across two funding rounds at a $2 billion valuation, with investors including Bezos Expeditions, Gates Frontier, Khosla Ventures, and NVIDIA's venture arm.
The origins of Field AI trace back to research conducted at NASA JPL between 2016 and 2022. Ali-akbar Agha-Mohammadi, who held the title of Technologist and Group Leader at JPL, served as Principal Investigator for several large-scale autonomy programs that would form the intellectual foundation of the company.
In 2016, the team began work on NASA's BRAILLE (Biologic and Resource Analog Investigations in Low Light Environments) project, which involved developing autonomous cave-exploration robots tested at Lava Beds National Monument in California as an analog for potential subsurface environments on Mars. The project required robots that could navigate without GPS or prior maps, a constraint that would define FieldAI's entire approach years later.
From 2018 to 2020, Agha-Mohammadi led Team CoSTAR (Collaborative SubTerranean Autonomous Robots), a joint team from JPL, MIT, Caltech, KAIST, and Lulea University of Technology, in the DARPA Subterranean Challenge. The competition challenged teams to build robots capable of searching underground mines, caves, and urban tunnels for objects of interest while operating autonomously. Team CoSTAR won the Urban Circuit in 2020 by achieving what the team described as the first fully autonomous many-kilometer exploration by a quadrupedal robot, while coordinating fleets of up to 11 heterogeneous robots including legged platforms, wheeled rovers, and aerial drones without GPS or pre-loaded maps. The software framework developed for the challenge, called NeBula (Networked Belief-aware Perceptual Autonomy), incorporated probabilistic world modeling and multi-robot coordination capabilities that became the conceptual ancestors of Field AI's later commercial products.
From 2021 to 2022, the group participated in DARPA RACER (Robotic Autonomy in Complex Environments with Resiliency), a program focused on autonomous navigation of military-style vehicles across unstructured off-road terrain at speed. The team became the first to successfully complete all eight required courses on the initial attempt in 2022, demonstrating multi-kilometer autonomous navigation without GPS, prior maps, or predefined trails across challenging terrain.
In 2023, Agha-Mohammadi formally established Field AI in Irvine, California, to commercialize the autonomy technology developed during the JPL years. The founding team drew heavily from the CoSTAR and RACER programs. David Fan, who had served as chief technologist on both DARPA programs and held a PhD from Georgia Tech, joined as Chief Technologist. Shayegan Omidshafiei, who had a doctorate from MIT and spent over five years as a Research Scientist at Google and DeepMind specializing in large-scale AI model training and multi-agent reinforcement learning, joined as AI and Science Lead. Sebastian Scherer, an Associate Research Professor at Carnegie Mellon University's Robotics Institute and a globally recognized field robotics expert, joined as Director of Fieldable Embodied AI. Eric Krotkov, the former Chief Science Officer at Toyota Research Institute and a veteran DARPA program manager, joined to lead federal operations. Duncan McIntyre, who had served as CFO of GoStudent and raised $600 million in that role while previously growing Delivery Hero's revenue from $100 million to $7.5 billion across 50 countries, became the company's Chief Financial Officer.
Field AI raised $91 million in a first funding round in late 2024. That round included participation from Khosla Ventures, Gates Frontier (Bill Gates's family office), Samsung, and Canaan Partners.
On August 20, 2025, the company announced it had closed a second round of $314 million, bringing cumulative fundraising to $405 million. The second round was co-led by Bezos Expeditions (Jeff Bezos's personal investment vehicle), Prysm Capital, and Temasek, the Singaporean state investment company. Additional participants included BHP Ventures (the venture arm of the global mining company BHP), Emerson Collective, Intel Capital, and NVentures (NVIDIA's corporate venture arm). The combined fundraising established a company valuation approaching $2 billion.
The second round was described as oversubscribed, which the company attributed to rapid customer adoption and multiple expansion contracts across hundreds of industrial deployment sites. The funds were designated for global expansion, doubling headcount by year-end, and further product development in locomotion and manipulation.
Vinod Khosla, the founder of Khosla Ventures, commented on the investment: "Enabling autonomy solutions at scale is an extremely difficult problem, but the deep expertise of the FieldAI team and their unique approach to embodied intelligence reflects a pragmatic path forward."
| Round | Date | Amount | Lead Investors |
|---|---|---|---|
| Seed/Series A | Late 2024 | $91 million | Khosla Ventures, Gates Frontier, Samsung, Canaan Partners |
| Series B | August 2025 | $314 million | Bezos Expeditions, Prysm Capital, Temasek |
| Total | $405 million |
Field AI's primary technical contribution is a class of AI models it calls Field Foundation Models (FFMs). The term refers to AI systems that were built from the ground up for embodied robotic intelligence rather than adapted from large language models or vision-language models originally designed for text or image tasks.
CEO Ali Agha has articulated the rationale for this approach in several public statements. In the company's August 2025 press release, he stated: "Rather than attempting to shoehorn large language and vision models into robotics, only to address their hallucinations and limitations as an afterthought, we have designed intrinsically risk-aware architectures from the ground up."
The central concept in the FFM architecture is a Belief World Model (BWM), a probabilistic representation of the environment that the robot updates continuously as it gathers sensor data. Rather than classifying terrain features (slope, roughness, step height) as fixed values, the system expresses them as probability distributions, allowing it to quantify uncertainty and propagate that uncertainty into motion planning and decision-making. The approach draws directly from Bayesian estimation techniques refined during the NeBula/CoSTAR program, where robots had to maintain consistent world models across multi-kilometer explorations in GPS-denied underground environments.
The FFM architecture contains several distinct sub-models that operate together:
| Component | Function |
|---|---|
| Belief World Model (BWM) | Probabilistic environment representation updated from multimodal sensor streams |
| Dynamics Foundation Model (DFM) | Detects and responds to conditions such as slipping, stumbling, or unexpected terrain contact |
| Multiagent Foundation Model (MFM) | Enables fleets of heterogeneous robots to share environmental understanding and coordinate tasks |
| Safety Layer | Uses probabilistic confidence assessments to identify and reject dangerous action candidates |
Sensor inputs include cameras, LiDAR, radar, and inertial measurement units. All inference runs on-device at latency under 100 milliseconds, enabling offline operation in environments without network connectivity. This on-edge design is considered important for mining and underground construction applications where cellular or Wi-Fi coverage is absent.
The company published its first patent filing (US20250252306A1) in August 2025. The patent, listing inventors Samuel Triest, David Fan, and Ali Agha, addresses traversability estimation for autonomous navigation. The core technique involves generating synthetic point clouds to train uncertainty-aware terrain models, allowing robots to assess navigation confidence across novel surfaces in real time.
Although Field AI is a software company, the FFMs are designed to be embodiment-agnostic, meaning a single trained model can control quadrupeds, wheeled robots, tracked vehicles, humanoid robots, and passenger-scale autonomous vehicles. This stands in contrast to earlier robotics software that required separate control stacks for each hardware platform.
Field AI's commercial product is branded as EDGE, described as a general-purpose robot brain that runs the Field Foundation Models on customers' existing hardware. The EDGE platform includes the full software stack covering mobility, high-level mission planning, multi-robot coordination, and mission execution. Customers integrate EDGE with their robot fleets through an initial hardware integration engagement and then pay recurring software licensing fees.
The hardware-agnostic architecture means customers can retrofit existing robots rather than purchasing new equipment. This has proven relevant for construction companies and mining operators that already own Boston Dynamics Spot units or other industrial platforms and want to upgrade their autonomy capabilities without hardware replacement.
Field AI's collaboration with NVIDIA deepened in 2025 through integration of NVIDIA Omniverse NuRec, which converts raw sensor data gathered by deployed robots into interactive 3D digital environments. These environments feed into NVIDIA's Isaac Sim and Isaac Lab frameworks for further training and policy validation. The pipeline creates a continuous improvement cycle where real-world robot data from job sites in North America, Europe, and Asia feeds back into simulation, which produces updated models pushed back to deployed robots. According to Ali Agha, digital twin creation that previously required three and a half months now takes 12 hours through this pipeline, a roughly 200-fold reduction.
Construction represents Field AI's largest publicly visible deployment vertical. Construction sites are among the most challenging environments for autonomous robots: layouts change daily as new structures are erected, heavy machinery moves unpredictably, workers create dynamic obstacles, and surface conditions vary from concrete slabs to mud and debris.
In November 2025, Field AI published a case study describing its deployment with DPR Construction, a major American general contractor, on a data center building project. Boston Dynamics Spot quadrupeds equipped with the EDGE platform operated autonomously across the site, capturing documentation that would previously have required engineers to walk manually. Over the course of the deployment, the robots traversed more than 100 miles on the job site, captured more than 45,000 photographs, mapped four complete floors, and documented 125,000 square feet of roofing and 500,000 square feet of interior spaces. The system generated a continuous digital record that project managers accessed through a mission dashboard, with deviation analysis and risk flagging processed on-device.
Justin Schreiner, a Senior Superintendent at DPR, described the impact: "The FieldAI system makes us better at what we do. Giving us greater efficiency, helping us document items more effectively, and taking some of the more mundane tasks off our plate, letting us focus on the more detailed, critical tasks."
Field AI and Boston Dynamics formalized their relationship through a partnership announced in March 2026, combining Boston Dynamics's Spot platform with Field AI's FFMs for uncharted exploration. Boston Dynamics founder Marc Raibert stated: "Combining Field AI's expertise in risk-aware autonomy with Spot's mobility allows us to tackle uncharted environments together." The partnership is creating one of the largest third-party quadruped fleets in the world, with Spot units deployed at construction sites across Asia, Europe, and North America. Compared to manual inspection processes, the partnership has demonstrated a 90% reduction in inspection and documentation time.
BHP Ventures' participation in Field AI's August 2025 funding round reflects the mining industry's interest in the company's GPS-denied autonomy capabilities. Underground mines and large open-pit operations share several characteristics with the SubT Challenge environments where the CoSTAR team originally developed NeBula: absent GPS, poor lighting, dust, dynamic ground conditions, and the need for robots to operate over many kilometers without constant human supervision.
Field AI's Dynamics Foundation Model is particularly relevant for mining and energy applications, where robots may traverse extremely rough terrain, steep slopes, or wet surfaces where wheel slip and unexpected contacts are common. By modeling these conditions probabilistically rather than as binary pass/fail classifications, the system can continue operating safely in marginal conditions rather than halting prematurely.
The company's pipeline inspection capability is relevant to energy infrastructure. Long-distance inspection of pipelines involves traversals of tens of kilometers across varying terrain, an environment where GPS dropout can occur and pre-mapped data becomes stale as terrain changes seasonally.
Field AI maintains a Federal division led by Eric Krotkov, targeting defense and government applications of the same GPS-denied autonomy platform used in commercial settings. The company's website lists federal applications including reconnaissance in contested environments, autonomous logistics, and inspection of critical infrastructure.
The company's DARPA lineage is a relevant credential for federal contracting. Ali Agha's role as Principal Investigator on two major DARPA programs and Krotkov's background as a DARPA program manager and leader of three DARPA Grand Challenges give the company familiarity with government acquisition processes and requirements.
Field AI has disclosed deployments in urban delivery and inspection, though specific customer names in these verticals have not been made public. The company's Multiagent Foundation Model enables fleets of robots to coordinate across urban environments, sharing sensor data and dividing coverage areas without requiring centralized coordination infrastructure.
| Partner | Nature of Partnership | Year Announced |
|---|---|---|
| Boston Dynamics | Integration of EDGE platform with Spot quadruped; joint deployment at construction sites | 2026 |
| NVIDIA | Integration with Omniverse NuRec, Isaac Sim, and NVIDIA Physical AI Data Factory Blueprint | 2025 |
| DPR Construction | Customer deployment for autonomous site documentation on data center projects | 2025 |
| BHP Ventures | Strategic investor with implied focus on mining autonomy applications | 2025 |
Field AI competes in a growing segment of companies developing general-purpose AI for robotics, often grouped under the term "embodied AI" or "physical AI." The three most frequently discussed companies in this segment are Field AI, Skild AI, and Physical Intelligence.
| Dimension | Field AI | Skild AI | Physical Intelligence |
|---|---|---|---|
| Founded | 2023 | 2023 | 2024 |
| Headquarters | Irvine, CA | Pittsburgh, PA | San Francisco, CA |
| Total funding | $405 million | $1.7 billion+ | $1 billion+ |
| Valuation | ~$2 billion | ~$14 billion (reported) | ~$5.6 billion |
| Primary environment | Outdoor, unstructured, GPS-denied industrial | Indoor and outdoor, general purpose | Indoor, dexterous manipulation |
| Core approach | Physics-first, risk-aware FFMs with probabilistic world modeling | Omni-bodied brain trained on human video and simulation | Vision-language-action models for dexterous manipulation |
| Hardware targets | Quadrupeds, wheeled robots, tracked vehicles, humanoids, passenger-scale vehicles | Quadrupeds, humanoids, tabletop arms, mobile manipulators | Manipulator arms, mobile manipulators, humanoids |
| Primary verticals | Construction, mining, energy, federal, logistics | Manufacturing, warehousing, data centers, construction | Home environments, light manufacturing |
| GPS dependence | Designed for GPS-denied operation | Not specifically focused on GPS-denied environments | Indoor focused, GPS not primary consideration |
| Founder background | NASA JPL, DARPA programs | Carnegie Mellon robotics research | Academic research, language modeling |
Field AI's most distinctive positioning relative to Skild AI and Physical Intelligence is its emphasis on outdoor, GPS-denied environments and its explicit commitment to physics-based uncertainty modeling. Skild AI, founded by Deepak Pathak and Abhinav Gupta from Carnegie Mellon, trains its Skild Brain on internet video of humans performing tasks and in physics simulation, with a goal of generalizing across robot embodiments in controlled or semi-controlled environments. Physical Intelligence, founded by Sergey Levine and colleagues, focuses primarily on dexterous manipulation tasks in indoor settings, with its pi0 model achieving approximately 80% task success in unseen home environments. Neither company has announced mining or large-scale outdoor construction deployments comparable to Field AI's publicly described work.
The funding gap between Field AI ($2 billion valuation) and Skild AI (reported $14 billion valuation) and Physical Intelligence ($5.6 billion valuation) reflects partly differing investor assessments of market size, partly differences in how far along each company is in scaling revenue, and partly the specificity of Field AI's industrial focus versus the more general positioning of its competitors.
Field AI's claims about its Field Foundation Models rest on real-world deployments rather than published benchmarks against standard robotic navigation datasets, which makes independent technical evaluation of the system's capabilities difficult. The company has not published extensive peer-reviewed literature comparing FFM performance to competing approaches in controlled conditions.
The hardware-agnostic architecture introduces integration complexity. Each new robot platform requires a hardware integration engagement before the EDGE software can run reliably, and the quality of the integration depends on the quality of the hardware's sensor suite. Robots with poor sensor configurations may not provide sufficient input quality for the probabilistic models to perform well.
Original equipment manufacturers (OEMs) of robotic platforms are increasingly developing their own proprietary autonomy software as a competitive differentiator. Boston Dynamics, for example, already ships Spot with built-in autonomy software. As OEMs build more capable native autonomy, the competitive rationale for a third-party software layer like EDGE may narrow, particularly for customers satisfied with a single robot platform.
BHP Ventures' investment is a strong signal of interest from the mining industry, but large-scale commercial mining deployment of autonomous robots in genuinely GPS-denied underground environments involves regulatory, safety, and operational complexity beyond what has been demonstrated in construction settings. No specific mining production deployment has been publicly confirmed by Field AI as of mid-2026.
The company remains private with no published revenue figures. Sacra estimated the company's business model as combining upfront hardware integration fees with recurring software licensing, which is a model that requires sustained customer retention to justify the valuation.