Robotics foundation models and embodied AI

1. Market thesis

Robotics is being restructured from bespoke, task-specific automation toward general-purpose embodied systems, with value shifting from individual robot hardware toward proprietary real-world data, foundation-model policies, simulation infrastructure, and fleet-scale deployment.

2. Market map

Segment 1: Robotics foundation models and policy software

Companies:

• Physical Intelligence — develops general-purpose models intended to control different robot embodiments and perform multiple manipulation tasks.
• Skild AI — building a general-purpose “robot brain” trained across tasks, environments, and robot platforms.
• Covariant — develops AI models for warehouse manipulation and has marketed a robotics foundation model, RFM-1.
• Google DeepMind — develops robotics models including RT-2 and Gemini Robotics-related capabilities.
• NVIDIA — provides Isaac foundation-model tooling, simulation, perception, and inference infrastructure rather than operating a single robot fleet.
• Intrinsic — Alphabet-backed robotics software company focused on general-purpose industrial robot intelligence and skill learning.

Dynamics: The segment is highly strategic but still early: model performance is improving rapidly, while defensibility depends less on model architecture than on proprietary robot-interaction data, deployment feedback loops, and the ability to support many hardware configurations.

---

Segment 2: Robot data, simulation, and development infrastructure

Companies:

• NVIDIA — Isaac Sim, Isaac Lab, Omniverse, Jetson, and related tooling form one of the most complete simulation-to-deployment stacks.
• Siemens — industrial digital twins, automation software, and simulation capabilities through its industrial software portfolio.
• Unity — real-time 3D and simulation tools used in robotics and synthetic-data workflows.
• Parallel Domain — synthetic-data and simulation platform best known for autonomous vehicles but applicable to embodied-AI training.
• Scale AI — data annotation, evaluation, and model-development infrastructure, including work relevant to robotics and autonomous systems.
• Dataloop — data management, annotation, and human-in-the-loop tooling for AI teams, including robotics use cases.

Dynamics: Simulation and data tooling are becoming the “picks and shovels” of the market, but pure synthetic data is unlikely to replace physical-world data; the strongest platforms will connect simulation, teleoperation, annotation, evaluation, and fleet learning.

Adjacent/less certain: Unity, Parallel Domain, and Dataloop have credible relevance to robotics, but robotics is not their sole or necessarily dominant market.

---

Segment 3: General-purpose robot OEMs and embodied platforms

Companies:

• Figure AI — humanoid robots aimed initially at industrial and logistics environments; developing its own embodied-AI stack.
• Agility Robotics — develops the Digit humanoid robot for logistics and industrial tasks.
• Apptronik — develops Apollo, a general-purpose humanoid robot for industrial applications.
• 1X — develops humanoid and wheeled robots, with emphasis on learning from real-world operation.
• Sanctuary AI — develops humanoid robots and a general-purpose control system known as Carbon.
• Tesla — developing Optimus and leveraging its broader autonomy, perception, and manufacturing capabilities.

Dynamics: Hardware remains capital-intensive and operationally difficult, but robot OEMs with manufacturing scale, captive deployment data, and strong software control may capture substantial value; the market is likely to consolidate around a few platforms rather than dozens of enduring humanoid brands.

---

Segment 4: Warehouse, industrial, and service deployment platforms

Companies:

• Amazon Robotics — large-scale deployment of mobile robots, robotic arms, and warehouse automation inside Amazon’s fulfillment network.
• Symbotic — automated warehouse systems combining software, storage, mobile robots, and robotic handling.
• Ocado — highly automated grocery-fulfillment systems using coordinated robots and proprietary orchestration software.
• Dexterity — develops AI-enabled robotic systems for palletizing, truck loading, and warehouse manipulation.
• Teradyne — owns Universal Robots and Mobile Industrial Robots, giving it a major position in collaborative and mobile industrial robotics.
• ABB — global industrial robot, automation, and software provider with broad installed-base access.

Dynamics: This is currently the clearest path to revenue because buyers pay for throughput, labor substitution, uptime, and integration—not for a model in isolation; deployment companies also generate the real-world data that foundation-model providers need.

---

3. Which segment captures the most value?

Likely winner: Deployment platforms and robot OEMs with proprietary fleets

The highest near- to medium-term value is likely to accrue to companies that control both the robot and the operating environment, rather than to standalone model vendors.

Why:

1. They own the customer relationship and budget. Buyers purchase warehouse throughput, manufacturing output, or labor savings—not an abstract policy model.
2. They capture recurring operational revenue. Robotics-as-a-service, maintenance, software subscriptions, and fleet management can be more valuable than one-time hardware sales.
3. They generate proprietary physical-world data. Every deployed robot creates demonstrations, failure cases, and environment-specific data that improve the system.
4. They can bundle the stack. Hardware, perception, control, simulation, deployment, and service contracts create switching costs.
5. Foundation models may commoditize model access. As open-source and hyperscaler models improve, a standalone model provider may face pricing pressure unless it has unique data or a differentiated deployment channel.

The important caveat is that foundation-model companies could become the long-term winner if one achieves a broadly transferable policy layer that becomes the default operating system across many robot manufacturers. In that scenario, the model provider could take software-like margins across a very large installed base.

4. White-space opportunities

Opportunity 1: Cross-embodiment fleet-learning and evaluation infrastructure

There is no obvious neutral platform that lets manufacturers and operators:

• pool and normalize data across different robot types;
• convert teleoperation, video, force, and proprioceptive data into reusable training datasets;
• evaluate policies consistently across simulation and physical environments;
• manage privacy, customer data rights, and model updates;
• deploy one policy across heterogeneous fleets.

Most current players are vertically integrated, hardware-specific, or focused on annotation/simulation alone. A trusted “GitHub plus observability layer for physical-world robot behavior” could become core infrastructure.

Opportunity 2: Robotics deployment middleware for brownfield industrial environments

A major gap exists between impressive demonstrations and deployment in older factories, warehouses, hospitals, and commercial buildings. A platform could provide:

• automatic site mapping and workflow discovery;
• integration with legacy PLCs, WMS, MES, ERP, and building systems;
• robot-agnostic task orchestration;
• safety certification and audit trails;
• human-robot exception handling;
• continuous ROI measurement and fleet optimization.

Current companies generally specialize in a robot, a vertical workflow, or a systems-integration project. There is still room for a neutral, model-agnostic “deployment operating system” for heterogeneous robots operating in existing facilities.