Humanoid Robots Reach Production Scale: Robotics Summit Opens on ROS vs. Proprietary Physical AI

The 2026 Robotics Summit & Expo opens Wednesday at the Thomas M. Menino Convention and Exhibition Center with a question that now carries real money behind it: who owns the software foundation for an industry whose hardware has started shipping?

The shift from "humanoids might someday work" to "humanoids are working" happened in the past 18 months. Figure AI's Figure 02 robot completed an 11-month deployment at BMW's Spartanburg, South Carolina plant, contributing to the production of more than 30,000 BMW X3 vehicles and handling more than 90,000 sheet-metal components across roughly 1,250 operating hours. Agility Robotics' Digit has moved more than 100,000 totes in live commercial operation at a GXO Logistics facility in Flowery Branch, Georgia, marking the industry's first formal commercial humanoid deployment under a Robots-as-a-Service model. Boston Dynamics' electric Atlas has begun commercial shipments, with its entire 2026 production allocation committed to Hyundai and Google DeepMind. The question no longer is whether humanoid robots can work. The question is what software will run them — and who controls that software.

ROS Faces Defining Test Against Proprietary Platforms

Open Robotics, the organization that maintains the Robot Operating System (ROS) and the Gazebo simulation platform, will open the summit with a keynote titled "An Open Foundation for the Age of AI-Powered Robots." The presenter is Brian Gerkey, co-founder of Open Robotics and currently CTO at Intrinsic — Google's robotics subsidiary — as well as board chair of the Open Source Robotics Alliance (OSRA), the governance body modeled on the Linux Foundation. Gerkey will make the case that open-source infrastructure is not merely a development convenience — it is the essential scaffolding for collaborative, trustworthy physical AI development at scale.

His timing is deliberate and his adversary is clearly named, if not directly. At CES 2026 in January, NVIDIA unveiled a full-stack ecosystem for physical AI — open foundation models on Hugging Face, simulation tooling, and edge hardware — positioning itself as what CEO Jensen Huang called the "ChatGPT moment for robotics." The company's stated ambition is to become the default platform for general-purpose robots, much as Android became the operating system for smartphones. Google DeepMind's Gemini Robotics platform, Physical Intelligence's π0 model, and Covariant's robot learning frameworks are each advancing proprietary model stacks on parallel tracks. The question of which layer of the stack remains open — and which becomes a commercial dependency — carries structural consequences for the robotics industry that will outlast any single product announcement.

ROS runs on millions of deployed robots worldwide and is the default starting point for most commercial robotics development. Its next test is whether it can absorb a fundamental architectural shift: the rise of vision-language-action (VLA) models, which unify perception, natural language understanding, and physical control into a single system. Where earlier robot software required separate models for each sensor modality and task, VLA models allow a robot to interpret what it sees, understand a language instruction, and execute movement without retraining for each new environment or object. Figure AI's Helix, NVIDIA's GR00T N1, and Google DeepMind's Gemini Robotics are all VLA implementations in active use. Whether ROS evolves as a native substrate for these models — or becomes a legacy layer underneath proprietary platforms — will, in large part, determine whether the industry's software layer consolidates around a corporate vendor or remains accessible to the broader engineering community.

Humanoid Robot Factory Deployments Drive Logistics Track

The summit's logistics automation track reflects where the economic argument for humanoid deployment is currently strongest. Unitree's IPO prospectus, filed on Shanghai's STAR Market in March 2026, disclosed that the average selling price of its humanoid robots fell from approximately $85,000 in 2023 to approximately $25,000 in 2025 — a drop of more than 70 percent in two years — while gross margins simultaneously improved to nearly 60 percent. Broader market data shows the trend is not isolated to Unitree: costs across the sector have fallen 40 to 60 percent as production scales and designs mature.

Session contributors at the summit include case study presenters from Agility Robotics — majority-owned by Amazon, with Digit units operating commercially at GXO Logistics facilities — and Figure AI, which completed its BMW Spartanburg trial with Figure 02 and is now deploying its next-generation Figure 03 platform. Toyota Motor Manufacturing Canada signed a commercial Robots-as-a-Service agreement with Agility Robotics in February 2026, deploying Digit humanoids at its Woodstock, Ontario plant. Mind Robotics, founded in November 2025 as a Rivian spinout and now valued at $2 billion after a $500 million Series A in March 2026, is also on the program. Mind Robotics explicitly focuses on non-humanoid industrial robots — founder RJ Scaringe has said "doing cartwheels does not create value in manufacturing" — and its inclusion signals the summit's breadth beyond bipedal platforms.

Not every deployment story is uncomplicated. Unitree's robots, which operate in Chinese EV factories and have been tested in research institutions globally, are subject to China's National Intelligence Law, which requires Chinese companies to cooperate with state intelligence agencies on demand regardless of stated privacy policies. A bipartisan congressional letter to the Defense Secretary, Commerce Secretary, and FCC Chairman, and a House Select Committee hearing on Chinese-linked robotics platforms, raised concerns about Unitree's "CloudSail" remote-access feature and documented connections to People's Liberation Army-affiliated institutions. The House Homeland Security Subcommittee on Cybersecurity and Infrastructure Protection held a dedicated hearing on Unitree's national security implications on March 17, 2026. A bill introduced May 21, 2026 — the "Blocking CCP Spy Tech Act of 2026" — would require a national security investigation of Unitree and, if found a risk, add it to the FCC's Covered List of equipment posing security threats. Companies evaluating Chinese-manufactured humanoid platforms should review these proceedings before procurement decisions.

How Does Vision-Language-Action AI Control Robots?

The term "VLA model" has become the industry's shorthand for the architectural leap that makes a robot programmable in natural language rather than explicit code. A VLA model takes a camera feed and a sentence — "pick up the blue cup and place it in the bin" — and outputs continuous motor commands directly, without a separate perception pipeline, a separate planning module, and a separate control system stitched together by hand. The architectural integration is consequential: modular systems fail at the boundaries between their components, particularly when environments change or instructions are ambiguous. VLA models handle variation better than any previous approach because reasoning and acting happen in the same learned system. The practical tradeoff is that failure recovery — knowing how to get back on track when a grasp slips — remains a weak point, as does running a VLA model large enough to reason well in real time directly on the robot itself without cloud support.

The practical implication for engineers at this summit is that selecting a VLA training infrastructure means, in effect, selecting a software dependency. NVIDIA's Isaac GR00T N1 model is open-weight but runs best on NVIDIA Jetson hardware. Google DeepMind's Gemini Robotics integrates with Google's broader AI cloud. Physical Intelligence's π0 model is proprietary. Open Robotics' argument — and the case Gerkey will make from the stage on Wednesday — is that a governance model like OSRA's can make ROS a trustworthy, enterprise-grade foundation rather than a fragmented collection of community repositories, keeping the core infrastructure of robotics development in the commons rather than inside any single vendor's platform.

Who Attends and What Takeaways Await

The two-day summit draws more than 6,000 robotics developers and engineers from aerospace and defense, healthcare, logistics, and manufacturing. Beyond the keynotes, the AI track includes a live on-stage build of an embodied AI robot by RealSense's VP of Developer Ecosystem, a panel on building the data flywheel for AI-native robots, and a session from Brain Corp's CTO on how new approaches in perception and continuous learning are redefining autonomy in retail and logistics environments. The summit is co-located with DeviceTalks Boston.

For attendees — and for the much larger community of engineers who will follow coverage — the practical output of the software debate playing out in the keynote and hallway conversations this week is not purely philosophical. Companies choosing a robotics software stack in 2026 are making a decision that will be costly to reverse in three years. The case for open-source infrastructure, as Gerkey will argue it, is that proprietary platforms introduce long-term vendor lock-in at precisely the moment when the hardware itself is becoming commoditized. The case for proprietary platforms, as NVIDIA and others have made it, is that vertically integrated tooling ships faster and works better out of the box. Both arguments are correct for different definitions of "better." That is the debate the Robotics Summit opens this week.

Frequently Asked Questions

Which humanoid robots are being deployed in factories in 2026?

Figure AI's Figure 02 completed an 11-month deployment at BMW's Spartanburg plant, contributing to production of more than 30,000 vehicles, before retirement in favor of the Figure 03. Agility Robotics' Digit is commercially deployed at GXO Logistics facilities under a Robots-as-a-Service model and at Toyota Motor Manufacturing Canada's Ontario plant. Boston Dynamics' Atlas has begun shipments to Hyundai's Robotics Metaplant Application Center and to Google DeepMind, with its entire 2026 production allocation already committed.

What is the Robot Operating System, and why does the open-source question matter?

The Robot Operating System (ROS) is an open-source software framework that provides the communication middleware, simulation tools, and reusable libraries that power most commercially deployed robots worldwide. The open-source question matters now because the industry is choosing whether its AI training infrastructure — vision-language-action models and the data pipelines that feed them — will be governed as a shared commons or owned by individual platform vendors like NVIDIA or Google DeepMind. That choice shapes who can build robots affordably, and on whose terms.

What companies are deploying humanoid robots in warehouse automation?

Agility Robotics is the furthest along in warehouse deployment, with Digit moving more than 100,000 totes at a GXO Logistics facility in Georgia and expanding to Toyota's Ontario auto plant. Boston Dynamics' Stretch robot handles box movement at logistics sites. Figure AI's Figure 03 is targeting warehouse and manufacturing environments. Unitree's robots operate in Chinese EV factory environments, though the company faces active U.S. congressional scrutiny over national security concerns related to China's National Intelligence Law.

What is physical AI, and how does it differ from conventional software AI?

Physical AI refers to AI systems that perceive, reason about, and act in the physical world through robotic bodies, rather than processing text or images for output on screens or through speakers. The technical distinction is that physical AI models must produce continuous motor commands in real time rather than discrete text tokens, and the consequences of errors are physical rather than informational. Vision-language-action models are the current leading approach to physical AI, combining visual perception, language understanding, and motor control in a single end-to-end system.