The Infrastructure Layer: NVIDIA

Before covering the robot makers, it's worth understanding NVIDIA's role — because it's not building robots, it's building the infrastructure that all the robot makers depend on. NVIDIA's goal is explicit: become "the Android of robotics" — the platform that all physical AI systems run on, regardless of who builds the hardware.

The Isaac GR00T platform provides a foundation model trained on thousands of hours of teleoperation data across multiple robot bodies. N1.6, released at CES 2026, is a 32-layer diffusion transformer that robot manufacturers can use as a starting point rather than training intelligence from scratch. Paired with the Isaac simulator — which lets companies train robots in virtual environments before deploying them physically — NVIDIA's strategy is to become indispensable to every company on this list.

For context: NVIDIA's position in embodied AI is analogous to its position in the disembodied AI revolution. Just as GPUs became the required infrastructure for training large language models like Claude and ChatGPT, NVIDIA's robotics infrastructure is being positioned as required substrate for physical intelligence.

Tesla — Optimus: Scale as the Strategy

Tesla's Optimus humanoid robot is perhaps the highest-profile entrant in the space, backed by Elon Musk's persistent attention and Tesla's manufacturing infrastructure. The strategic logic is straightforward: Tesla already builds sophisticated hardware at scale (cars), already trains complex AI systems (Autopilot/FSD), and already operates Gigafactories that could serve as both production facilities and real-world training environments for Optimus.

Optimus units are currently deployed inside Tesla's own Gigafactories, performing manufacturing tasks. This is both a deployment and a data collection strategy — every hour of Optimus operation in a factory generates training data that improves the system. Tesla aims for millions of Optimus units annually by the 2030s, with a target price of $20,000-$30,000 per unit at scale.

The criticism: Tesla has a long history of optimistic timelines. Autopilot's full self-driving capability has been "one year away" for roughly a decade. Optimus is impressive in demonstration; whether it can achieve the reliability required for safe, unsupervised production deployment on the timelines Tesla projects remains genuinely uncertain.

Figure AI — Moving Fast with Enterprise Partners

Figure AI has made the fastest real-world deployment progress of any pure-play humanoid robotics startup. Its Figure 02 robot completed 1,250 operational hours on an active BMW automotive production line — not a demonstration, but actual production work validated by one of the world's most demanding manufacturers.

Figure's approach is to partner deeply with major industrial customers who provide both revenue and real-world training environments. BMW, Mercedes-Benz, and others have taken equity stakes in robotics companies specifically to secure supply — a signal that strategic industrial buyers believe the technology is close enough to production-relevance to warrant investment now.

The company has also partnered with OpenAI — whose language model reasoning capabilities power Figure's instruction understanding, allowing operators to give natural language commands that the robot interprets and executes.

Boston Dynamics — The Track Record Leader

Boston Dynamics, now a subsidiary of Hyundai Motor Group, has the longest track record of any company on this list. Its robots — Spot, Stretch, and the humanoid Atlas — have been operational in real-world environments for years, giving it a reliability data advantage that newer entrants simply don't have.

In 2026, Boston Dynamics announced production-ready Atlas units for industrial deployment, focusing on manufacturing and construction use cases where Atlas's strength and agility provide advantages that purpose-built industrial robots can't match. Spot continues to be deployed in hazardous environment inspection — oil and gas facilities, construction sites, mines — where its quadruped design excels.

Boston Dynamics' Orbit software adds the enterprise AI layer — fleet management, data analytics, and integration with existing industrial systems — positioning it as a complete platform rather than just a hardware provider.

Agility Robotics — Purpose-Built for Logistics

Agility Robotics' Digit is purpose-built for a specific use case: operating alongside humans in logistics and warehouse environments. Unlike humanoids designed for general-purpose tasks, Digit is optimized for the repetitive, high-volume material handling tasks that define warehouse work — picking items, moving totes, unloading trucks.

Amazon has been a key partner and investor, testing Digit in its facilities as part of its broader strategy to automate the parts of its warehouse operations that aren't yet handled by its existing robot fleet. Agility's RoboFab facility has a production capacity of 10,000 Digit units annually — modest by automotive standards, but significant for the humanoid robotics industry.

Physical Intelligence — The Research-to-Reality Bridge

Physical Intelligence (pi) is one of the most technically ambitious companies in the space. Founded by former Google and Stanford researchers, it's focused specifically on the AI intelligence layer — developing foundation models for physical AI that work across diverse robot bodies and task types.

Its pi-0.5 model demonstrated meaningful open-world generalization — the ability to perform tasks in environments and with objects the system hasn't been explicitly trained on. This is the key capability gap that separates current robotics from truly general-purpose physical AI, and Physical Intelligence is arguably further along on this specific problem than any other organization.

Physical Intelligence raised $400 million in 2024 and attracted significant talent from the top AI research institutions. It doesn't manufacture robots — it provides the AI foundation that robot manufacturers build on, similar to how OpenAI's models power other companies' products.

The Chinese Contenders: AgiBot, Unitree, UBTECH

The global embodied AI landscape is split in a way that mirrors the broader geopolitics of AI: China dominates in deployment volume, the US leads in AI software capability. China has 2 million factory robots versus 394,000 in the US — a 5x deployment gap — and is producing thousands of humanoid units through domestic manufacturers while the US market is still ramping.

AgiBot currently leads global humanoid shipments. Backed by Alibaba and benefiting from strong government support, it has been able to scale production faster than Western competitors, though at capability levels that industry observers consider behind the frontier AI approaches of Physical Intelligence and Figure.

Unitree Robotics has made waves with aggressive pricing — its Go2 and H1 models are available at price points that make them accessible to research labs and smaller companies, dramatically expanding the pool of organizations that can work with humanoid hardware. Unitree is filing for a $7 billion IPO with reported 60% gross margins — an unusually healthy margin for a hardware company.

UBTECH has deployed its Walker S2 at scale and has been testing multi-robot coordination systems that enable decentralized swarm intelligence — multiple robots sharing spatial awareness and collaborating on tasks that no single unit could handle alone.

The Infrastructure Enablers: Auki Labs and Spatial AI

One of the most interesting companies in the embodied AI ecosystem isn't building robots at all — it's building the spatial intelligence layer that robots need to understand physical environments. Auki Labs, based in Hong Kong, has developed what it calls the "posemesh": a decentralized network that translates physical environments into precise 3D digital coordinates that robots, AR devices, and AI systems can share and act on.

The practical implication is significant: rather than each robot spending time mapping a new environment from scratch, a robot entering a space can download a pre-built spatial map — understanding the layout, the locations of objects, and the navigable paths before taking a single step. It's the difference between a delivery driver memorizing a new city versus downloading Google Maps. We explore this concept in depth in our dedicated piece: Spatial AI and the Physical World: How Auki, Digital Twins, and Robots Are Bridging the Gap.

The AI Brain Makers: Google DeepMind, OpenAI

Both of the companies behind the most widely-used disembodied AI tools — Gemini comes from Google DeepMind, while ChatGPT comes from OpenAI — are actively developing physical AI capabilities.

Google DeepMind's robotics division has produced some of the most influential research in the field, including the RT-2 model (which applies the same scaling laws that improved language models to robot learning) and contributions to the Open X-Embodiment dataset — a collaborative effort across 34 labs to create shared training data for robotic manipulation.

OpenAI has not been as publicly active in robotics but is understood to be investing in the space, and its partnership with Figure AI suggests a path where its language model reasoning becomes the intelligence layer for physical AI systems. The same models that power ChatGPT Plus subscribers' conversations may eventually power the decision-making of robots on factory floors.

What to Watch

The companies with the most defensible positions are those accumulating the most embodied training data through deployed fleets — because this data doesn't exist on the internet and can't be scraped. Every operational hour of a robot in the real world generates data that can improve the next generation. The companies deploying now, even at modest scale, are building data advantages that will compound.

The companies with the most interesting long-term potential are those working on open-world generalization — the ability to handle novel environments and tasks without specific training. Physical Intelligence, Google DeepMind, and NVIDIA GR00T are all making meaningful investments here.

For the full picture on what's holding back even the best-funded companies on this list, read our companion piece: The Biggest Challenges Facing Embodied AI Right Now.