Physical AI and the Evolution of Robotics Foundation Models

Physical AI solutions increasingly leverage robotics foundation models to facilitate advanced automation capabilities. This Research Highlight shares ABI Research’s analysis of the market, spanning robotics foundation model revenue forecasts, adoption timelines, and core ecosystem synergies.

• Robotics foundation models are moving Physical AI beyond basic automation by enabling robots to handle dexterous tasks that were previously difficult to automate, especially with deformable objects.
• Manufacturing and warehousing are the clearest early markets because they offer structured environments where reliability and repeatability are easier to achieve.
• The market is still immature, with most activity centered on R&D, but long-term growth depends on progress across models, compute, simulation, and edge semiconductors.
• Commercial success will rely on ecosystem coordination. Model developers, cloud platforms, training tools, and chip vendors all need to work together to make Physical AI scalable in real-world deployments.

A trend set by NVIDIA, Physical Artificial Intelligence (AI) solutions using robotics foundation models require compute in three places: on-device chipset, a local server, and a cloud platform. Early Physical AI Proofs of Concept (PoCs) have underscored the value of manipulating deformable objects via foundation models, something that previous industry attempts have failed to do. This means objects such as plastic bags, clothing, and food can be physically altered. The commercial appeal is clear: automating repetitive manual processes like packaging, dexterous assembly, cleaning, cooking, folding, and more.

The robotics foundation model market remains immature, with most activity confined to Research & Development (R&D). However, ABI Research projects significant progress in the 2030s across many industries. We forecast the Total Addressable Market (TAM) for robotics foundation models to reach US$150 billion by 2036. Collaborative Robots (cobots) lead on the form factor side of things, while manufacturing and warehousing lead in terms of sector growth. Humanoids are more aligned with the long-term promise of general-purpose autonomy.

Market Outlook

Manufacturing and Warehousing Will Lead Robotics Foundation Model Adoption

Manufacturing and warehousing offer the clearest near-term opportunities for robotics foundation models. These work environments are highly structured, conducive to repeatability and reliability.

Industrial automation leaders such as ABB and Siemens are already pushing the technology forward, with manufacturing representing a potential US$30 billion TAM by 2036. Adoption among Small and Medium Enterprises (SMEs) will be more gradual, due to limited data infrastructure and a shortage of systems integrators.

Warehousing and logistics will represent another US$21 billion TAM by 2036. The sector remains highly manual, leaving plenty of repetitive and dexterous tasks for Physical AI to automate.

Robotics Foundation Models Will Move Beyond the Factory and Warehouse

Less structured environments present a tougher challenge for Physical AI, but lucrative opportunities are on the horizon. Healthcare applications are forecast to represent a US$16 billion TAM by 2036, as hospitals automate administrative and support tasks. Restaurants, retail, and hospitality could contribute another combined US$27 billion. Here, foundation model-supported robots will automate back-of-house tasks.

Defense is also emerging as an early market for robotics foundation models; contractors such as Palladyne AI, Helsing, Anduril, and Palantir are developing automation solutions. Further out, industries such as oil & gas, mining, construction, and agriculture will benefit as foundation models get better at handling unpredictable environments and edge cases.

The automotive market will play a key role in expanding robotics foundation models to more dynamic enterprise settings. Autonomous Driver-Assistance Systems (ADAS) within automotive will lead the way, imparting important lessons on safety and legal liability for semi-structured environments.

Technological Underpinning of Physical AI and Robotics Foundation Models

The Physical AI and robotics foundation model landscape will not evolve without a unified effort across various technology fronts. AI models, compute, and semiconductor innovation will dictate market progression. ABI Research emphasizes the importance of how technology suppliers within these groups interact and address data, interoperability, and deployment challenges. The role of these companies is outlined below.

VLA Models

Vision-Language-Action (VLA) models remain the most visible part of the robotics foundation model landscape. Developers target a wide range of robot form factors, from industrial systems to humanoids. Leading players include Physical Intelligence, NVIDIA, Figure, Microsoft, OpenAI, Google DeepMind, 1X, Agility Robotics, Sereact, FieldAI, and Skild AI. While interest and investment remain high, VLA models are still at an early stage; commercial maturity is constrained by the non-deterministic nature of transformer-based systems.

Training, Orchestration, and Cloud Service Providers

Training and lifecycle support are becoming a core part of the Physical AI market as robotics foundation models require reinforcement learning, simulation, retraining, and virtual test environments. Cloud service providers AWS and Microsoft Azure are notable providers in this layer, while NVIDIA’s Isaac portfolio has become an important commercial training and deployment pipeline. Open-source toolchains, including LeRobot, MuJoCo, ROS 2, PyTorch/JAX, robomimic, and diffusion policy, also remain widely used in Research and Development (R&D) settings.

World Models

World models and related world action models are emerging as a key area of development for general-purpose robotics. Using real-world time-series data, these models “imagine” their environment and form predictive internal rollouts of it. A primary application will be where longer-horizon prediction and planning are needed. Important research efforts in this area include DeepMind’s Genie, Wayve’s embodied world models, Tesla’s neural simulation approaches, NVIDIA Cosmos, Physical Intelligence’s models, and adjacent video generation systems such as Sora. Reasoning-based approaches such as MolmoAct are also contributing to technical evolution, although they remain highly experimental.

Semiconductor Vendors

Semiconductor vendors provide the edge intelligence needed to process robotics-related data. NVIDIA Jetson currently dominates the silicon market in this context, but competition is expanding as more semiconductor vendors target robotics foundation model workloads. Intel, AMD, Qualcomm, and Ambarella are each developing relevant hardware and software stacks. Examples include Intel Core Ultra, AMD Ryzen AI Embedded, Ambarella CV7, and Qualcomm Dragonwing.

Learn More

Although robotics foundation models are in their genesis phase, analysts anticipate them to be an increasingly essential technology enabling groundbreaking applications previously impossible. For a comprehensive analysis of how foundation models will evolve and where business opportunities will materialize, download ABI Research’s Physical AI: Robotics Foundation Models report.