North America's largest robotics trade show opens Monday in Chicago with a different question than it has asked in any of its previous 49 editions: not whether humanoid robots are ready, but how fast industry can absorb them now that they are. The Association for Advancing Automation's Automate 2026 runs June 22–25 at McCormick Place, drawing more than 50,000 attendees, over 1,000 exhibitors, and 450,000 square feet of automation technology — the largest edition in the show's 50-year history. For the first time, a dedicated Humanoid Robot Pavilion sponsored by NVIDIA anchors the show floor, and the dominant theme across the keynote agenda is physical AI: the practice of training robots through demonstration rather than explicit programming.

The show opens two days after production humanoids became a literal line item. Figure AI's BotQ facility crossed a rate of one Figure 03 robot per hour — a 24x throughput increase in under 120 days, with more than 350 units delivered. Boston Dynamics began commercial shipments of its electric Atlas, with its entire 2026 production run committed to Hyundai's Robotics Metaplant Application Center and to Google DeepMind. Agility Robotics' Digit is working commercial shifts at Toyota Motor Manufacturing Canada's Woodstock, Ontario plant under a Robots-as-a-Service agreement.

First NVIDIA-Sponsored Humanoid Pavilion Signals Compute Bet on Industrial Robots

The most visible signal of the shift is structural: for the first time at Automate, humanoid robots have their own dedicated pavilion on the show floor, sponsored by NVIDIA. More than 20 humanoid robot organizations from around the world will demonstrate their platforms there, with free access for all registered attendees. A theater stage inside the pavilion gives exhibitors time to present their technologies throughout the week. A co-located Humanoid Robot Forum — a separate paid conference running Tuesday and Wednesday afternoons, June 23–24 — goes deeper into the engineering challenges, bringing together leaders from Boston Dynamics, NEURA Robotics, NVIDIA, and Toyota Research Institute for sessions on real-world commercialization, safety standards, and the enabling technologies that still separate a compelling demo from a reliable production deployment.

"Humanoid robotics is generating a tremendous amount of interest across the automation ecosystem," Association for Advancing Automation president Jeff Burnstein said ahead of the show. "By bringing the Humanoid Robot Forum and a dedicated pavilion to Automate, we're giving attendees a practical opportunity to learn where the technology stands today, what challenges remain, and how humanoids may fit into real-world applications over time."

NVIDIA's pavilion sponsorship is itself editorial: the chipmaker's inference platforms and Isaac GR00T training stack have become the dominant computational backbone for robot AI development, and its pavilion investment signals to the industry that humanoids represent a meaningful share of its addressable inference market.

How Physical AI Works: From Demonstration to Deployment

The defining phrase of Automate 2026's agenda — "physical AI" — describes a specific engineering shift with direct consequences for how manufacturers can deploy automation going forward.

Traditional industrial robots are programmed explicitly: an engineer writes a movement script telling the robot exactly where to move, in what sequence, at what speed. That approach works for high-volume, highly repetitive tasks performed in fixed environments. It breaks down whenever an object is positioned slightly differently, a new product variant arrives, or an unstructured task requires contextual judgment. This is why, as Standard Bots CEO Evan Beard will argue in his Wednesday keynote, roughly 99 percent of real-world manufacturing tasks have resisted automation: the programming burden for each variation is prohibitive.

Physical AI addresses this through imitation learning — a training paradigm in which a robot acquires behaviors by observing human demonstrations rather than following hand-coded routines. A human performs a task via teleoperation using a VR controller, by physically guiding the robot arm through the motion, or by wearing a head-mounted camera and performing the task naturally, and the robot's AI model learns the underlying policy from that observation data. The practical framing, as A3 summarized Beard's position: physical AI moves robotics from "program every step" to "show the robot what to do."

The architecture that currently dominates this approach is the vision-language-action (VLA) model: a single end-to-end learned system that takes a camera feed and a natural-language instruction and outputs continuous motor commands — no separate perception pipeline, no separate planning module. Figure AI's Helix, NVIDIA's Isaac GR00T N1, and Google DeepMind's Gemini Robotics are all VLA implementations in live industrial use.

The central engineering constraint that Automate 2026 will not solve — but will spend four days debating — is the sim-to-real gap: the disconnect between how an AI policy performs in simulation (where it trains on millions of synthetic interactions) and how it performs on a real factory floor (where friction, material variation, and contact dynamics behave differently than any simulator perfectly replicates). NVIDIA's pipeline attempts to close that gap by combining Isaac Sim for high-fidelity physics rendering, Isaac Lab for GPU-parallel policy training at scale, Cosmos world foundation models for physically accurate synthetic training data, and the Newton physics engine for contact dynamics. The on-device inference module — the chip that runs the trained policy inside the robot itself — is NVIDIA's Jetson Thor, a separate product from the Vera Rubin data-center platform that trains the models upstream. Whether that pipeline narrows the sim-to-real gap enough for the 99-plus percent task reliability that production manufacturing requires remains the central open question in industrial robotics as of the show's opening day.

Deployments Already Running: What the Data Shows

Three deployments have crossed from pilot to commercial production in the months leading to Automate 2026, and each establishes a data point the industry will use to calibrate expectations.

Figure AI's BotQ production ramp is the clearest manufacturing milestone. The company's California facility grew from one Figure 03 unit per day to one per hour in under 120 days, supported by custom manufacturing software operating across more than 150 networked workstations, with end-of-line first-pass yields above 80 percent. The ramp matters not only as a supply signal but as a data-generation signal: each robot deployed collects real-world operational data that feeds back into the training pipeline, accelerating model improvement.

Boston Dynamics' Atlas shipments mark the commercial debut of the first enterprise-grade humanoid designed from the ground up for automotive-grade supply chains. The robot features 56 degrees of freedom, a 110-pound lift capacity, 360-degree torso rotation, and autonomous battery swapping for continuous operation. All 2026 units are committed to Hyundai's Robotics Metaplant Application Center and Google DeepMind, with additional customers planned for 2027.

The Agility Robotics–Toyota Canada agreement is the model-by-model case study the industry has been waiting for. After a year-long pilot, Toyota Motor Manufacturing Canada signed a commercial Robots-as-a-Service contract for seven Digit humanoids at its Woodstock, Ontario facility producing the RAV4 and RAV4 Hybrid. The RaaS model — in which manufacturers lease rather than purchase robots — converts what would otherwise be a capital expenditure into an operating expense, lowering the financial barrier to humanoid adoption for manufacturers who are uncertain about long-term utilization.

A critical constraint persists across all three deployments: humanoid robots in 2026 remain well-suited to structured, repetitive industrial tasks but cannot reliably generalize to complex, variable, or safety-critical environments. Industry researchers have framed the reliability threshold directly — a demo that works 70 percent of the time is not good enough for manufacturing; 99-plus percent is the required floor. Safety certification for fenceless humanoid operation — robots working alongside human workers without physical barriers — also remains an open gap, with standards organizations still developing the applicable framework.

Keynotes: The 99% Problem and What Physical AI Changes

The keynote program takes a deliberately grounded approach. Monday opens with a leadership roundtable featuring executives from FANUC America, Schneider Electric, Cognex, and Intrinsic examining how manufacturers are integrating AI into existing automation systems. Tuesday's session, from Siemens Digital Industries' Annemarie Breu and Chris Stevens, addresses the workforce dimension of automation adoption.

The most analytically pointed keynote belongs to Evan Beard, co-founder and CEO of Standard Bots, who takes the stage Wednesday with a session titled "99% of Tasks Still Can't Be Automated: How Physical AI Changes That." The session targets the premise driving every deployment discussion at the show: most real-world manufacturing tasks have resisted automation not because of hardware limitations but because of the programming burden that explicit rule-based systems impose on every task variation. Beard's argument — and the argument embedded in every VLA model deployment on the show floor — is that imitation learning flattens that burden: a manufacturer who can demonstrate a task rather than program it can expand automation into a far larger fraction of their operations.

A3 has framed physical AI as the conceptual bridge between current industrial robots and the more adaptable machines the industry needs. Whether the bridge holds under production conditions — not demonstration conditions — is what the forum sessions will interrogate.

How NVIDIA's Physical AI Stack Became the Industry Default

The largest implication of Automate 2026 that the show's promotional materials do not name is the one that will shape the industry for the next decade: NVIDIA's physical AI platform — Isaac Sim, Isaac Lab, Cosmos, Newton, and Jetson Thor — has become the de facto training infrastructure for commercial humanoid robotics, and the show floor is where that consolidation becomes publicly visible.

ABB Robotics, FANUC, KUKA, Universal Robots, and Doosan Robotics are all building on the same NVIDIA stack. The pavilion sponsorship is not philanthropy: NVIDIA is positioning itself as the platform layer for an industry whose scale is measured in inference compute demand per robot deployed, multiplied by the number of robots deployed across global manufacturing. The competitive variable among the robotics companies on the show floor is no longer which tools they use — it is how far each has advanced from integration to validated industrial application.

The structural question this creates for manufacturers at the show is not primarily "which humanoid robot should we buy?" but "which combination of training architecture, deployment model, and software platform will we be dependent on in five years?" A choice made in 2026 in favor of a specific physical AI stack is a dependency that will compound as models are fine-tuned on proprietary operational data. The open-source alternative — the Robot Operating System and the governance framework being advanced by the Open Source Robotics Alliance — will be competing for mind share on the same show floor, even without a dedicated pavilion sponsor.

Startup Challenge and Full Show Floor

While the Humanoid Pavilion commands the editorial spotlight, Automate 2026 covers the full breadth of automation: autonomous mobile robots, machine vision systems, motion control hardware, industrial software platforms, and AI-enabled sensing across the North and South halls of McCormick Place.

The Automate Startup Challenge — co-sponsored by NVIDIA and Microsoft and offering a $10,000 prize — gives ten finalists from early-stage robotics and automation companies a public pitch platform on opening day, June 22. More than 200 expert speakers across 140-plus conference sessions round out four days that, in the aggregate, ask a single question: whether the manufacturing industry is organizationally ready to absorb a technology transition that its hardware suppliers have clearly declared is already underway.

Frequently Asked Questions

What is physical AI and how does it differ from conventional industrial robotics?

Conventional industrial robots are programmed with explicit movement instructions for each task. Physical AI trains robots through demonstration — a human performs a task while the robot's AI model observes and learns the underlying policy. The dominant implementation is the vision-language-action (VLA) model, which takes a camera feed and a natural-language instruction and produces continuous motor commands without separate perception, planning, or control modules. The practical consequence is that a manufacturer can expand automation to new tasks without writing new code for each variation, by showing the robot what to do instead.

What humanoid robots are being deployed commercially in factories in 2026?

Three humanoid platforms have crossed from pilot to commercial operation ahead of Automate 2026. Agility Robotics' Digit is deployed at Toyota Motor Manufacturing Canada's Woodstock, Ontario plant under a Robots-as-a-Service contract. Figure AI's Figure 03 is being produced at a rate of one unit per hour at BotQ, with more than 350 units delivered to industrial customers. Boston Dynamics' electric Atlas has begun shipping, with its entire 2026 production committed to Hyundai's Robotics Metaplant Application Center and Google DeepMind.

What is the sim-to-real gap, and why does it matter for industrial deployment?

The sim-to-real gap is the difference between how an AI robot policy performs in simulation — where it trains on millions of synthetic interactions — and how it performs on a real factory floor. Real environments involve friction, material variation, and contact dynamics that simulators approximate but do not perfectly replicate, causing trained policies to fail in ways they did not fail in training. Closing this gap enough to reach the 99-plus percent task reliability that production manufacturing requires is the central engineering challenge in physical AI. NVIDIA's pipeline of Isaac Sim, Cosmos world foundation models, and the Newton physics engine represents the current state of the art in synthetic training data generation for this purpose.

What does NVIDIA's pavilion sponsorship at Automate 2026 actually mean for manufacturers?

It means NVIDIA has identified humanoid robotics as a significant growth market for its inference compute platforms. Every humanoid robot deployed at commercial scale requires inference hardware to run its AI model in real time — the Jetson Thor module for on-device computation, and data-center-scale GPUs for training and fine-tuning upstream. Multiple major robotics manufacturers are already building on the same NVIDIA stack, which means a manufacturer choosing a physical AI platform in 2026 is also, in effect, choosing an AI infrastructure dependency. Understanding that dependency — and evaluating alternatives including open-source ROS-based stacks — is part of any serious humanoid procurement decision.