The Next Humanoid Robotics Race: Safer Robots, Bigger Scale, More Markets

Agility Robotics has accumulated more than 65,000 hours of operation, reported more than $300 million in multi-year orders for its next-generation robot, and helped shape a forthcoming ISO safety standard covering humanoid robots. Its playbook reveals more about the industry’s future than any demo video.

The first lap of the humanoid race was run on video: robots walking, lifting, balancing and backflipping across keynote stages and investor decks. The next lap will be run in warehouses, and scored differently: safer robots, deployed at larger scale, across more markets.

Few people describe that shift more concretely than Jonathan Hurst, co-founder and chief robot officer of Agility Robotics, interviewed at Machina 2026. His starting point is an old saying in robotics research: simulations are doomed to succeed. The step from simulation to a live demonstration is enormous, and the step from demonstration to deployment is just as large again.

“When you deploy something, it isn’t just a few weeks to show that it can do a task. It has to do that task, and it has to do that with enough uptime and enough throughput to meet the metrics to actually provide a return on investment for the customer.”

Once a robot enters a workflow, its failure is no longer an engineering setback; it is the customer’s operational problem, measured in stopped lines and lost money. That single fact reorganizes everything else: the safety case, the business model, the data strategy and the pace of expansion.

The context makes Agility’s position clearer. NVIDIA is building the enabling layer for physical AI, from chips and edge compute to open AI models, open teleoperation and simulation frameworks , open world models, developer tools and safety frameworks. Agility Robotics shows what the deployment layer must then deliver: real use cases, uptime, throughput, worker trust, safety cases, service models and measurable ROI.

Together, they show where physical AI is heading. The market is moving from what a humanoid can demonstrate once to what a humanoid can do repeatedly, safely and profitably.

A decade of homework

Agility has spent more than a decade moving from research to revenue, from the Cassie research biped in 2015 to Digit, its human-centric worker robot, now in its fourth generation. Digit v4 is performing paid work today for Schaeffler, Toyota Motor Manufacturing Canada, GXO and Mercado Libre. Across deployment commitments at nine customer facilities, Digit has accumulated more than 65,000 hours of operation.

The shareholder register reads like a map of the ecosystem around the company: NVIDIA, Amazon, SoftBank Vision Fund 2, Schaeffler and Foxconn are among its strategic investors.

A two-to-three-year moat, written into standards

As Agility moved from proof of concept to deployment, it faced a critical problem: existing safety standards for conventional industrial robots and automated production equipment did not fully address dynamically stable humanoids operating alongside people.

Agility therefore developed its own hazard and risk assessment process, mitigation strategies and third-party validation approach. It also helped write ISO/CD 25785-1, the forthcoming safety standard for dynamically stable industrial mobile robots, including bipedal humanoids. The standard remains under development.

That homework is starting to pay compound interest. Digit v5, planned for commercial launch in 2026, is built around what Agility calls cooperative safety: the ability to operate in the same space as people without traditional physical safety barriers. Agility expects the generational improvements to be substantial: payload capacity rising by 40% to 23 kg (50 lb), vertical reach by 30% to 2.2 m (7.2 ft), operating availability of up to 22 hours per day, and changeable end-effectors that can support a broader range of industrial tasks. Agility is also NVIDIA’s launch partner for Halos for Robotics, integrating NVIDIA IGX Thor and Halos Core into Digit’s proprietary safe human-detection system.

Once published, the standard should give the industry a common safety reference. It will also reinforce Hurst’s central point: compliance cannot be bolted on after the robot has been built.

“You can’t retrofit a robot safety standard. You’re going to have to design and build it from the ground up.”

In humanoid robotics, safety is not a cost line. It is a moat, and the gateway to every new market.

Trust is an architecture

What does designed-in safety look like? Digit carries a separate, fully islanded supervisory computer with absolute authority to shut the robot down. One operating rule forbids raising the hands above the waist near people, guarding against a finger in someone’s eye. If the main controller breaks the rule, whether through a bug, a fault or a hack, the supervisor shuts the robot down. A chaperone with an emergency stop walks the robot between work areas, and a red e-stop button sits on its back.

The detail is telling: the same isolated design supports both functional safety and cybersecurity. Trust, in this world, is not a claim. It is architectures.

Selling output, not robots

Agility’s business model completes the logic. Customers can buy robots outright with a service agreement, but the signature offer is robot-as-a-service, typically under multi-year contracts.

“The contract is not for us to deliver a Digit robot. The contract is for us to move these bins at a certain rate, at a certain throughput, a certain uptime,” Hurst explains.

At Schaeffler, Digit pulls steel bins off pallets and feeds an industrial washer that processes a fixed number of bins per minute. The robot must match that rate. A dropped bin scatters parts that must be scrapped for contamination and stops the line, so the robot’s reliability has to be roughly on par with a good human worker.

The model transfers performance risk from customer to vendor, which is exactly why it works. It forces Agility to industrialize uptime, fleet management, charging infrastructure and field service through Agility Arc, its cloud platform for fleet orchestration, and it gives customers the single, simple assurance they actually want: the work gets done. Digit has now moved more than 100,000 totes in a commercial deployment, suggesting that the operating model is holding.

Agility reports more than $300 million in multi-year contracted orders for Digit v5, subject to the achievement of certain contractual milestones, supported by a growing pipeline of more than 30 customers.

Demand is a labor problem

The pull is not fascination with humanoids. It is that customers cannot hire. Deloitte estimates that 1.9 million U.S. manufacturing jobs could go unfilled by 2033, and tasks that are dull, dirty or dangerous are the hardest to staff. Hurst rejects the replacement framing:

“I would not say it’s an upgrade to humans. I would say it’s an augmentation to people.”

Acceptance still has to be earned. The robot must be perceived as, and actually be, a helper; industrial design matters; workers need to know what it can and cannot do. Hurst’s experience is that anxiety about humanoids drops sharply once people see Digit handling the tasks nobody wanted. Hurst points to the transformation of agriculture, where mechanization dramatically reduced the share of people working on farms without producing permanent mass unemployment. Autonomous trucking, meanwhile, has advanced more slowly than many originally predicted, even as driver shortages have deepened.

Data quality and the 10,000-hour flywheel

On data, Hurst inverts the industry’s usual arithmetic:

“The quality of data is much more valuable than the quantity of data.”

Cognitive AI, he argues, is becoming a utility that robot makers will buy the way they buy electricity. Physical AI is different: balance, whole-body control and manipulation are inseparable from the robot’s body, so Agility builds them in-house. Agility trains today in physics-based digital twins of small environments; world models will matter when robots must reason across an entire warehouse; and the sim-to-real gap in friction, contact and compliance means deployment remains the only real exam.

The hardest problem ahead is judgment. Today, Digit follows strict rules when a person approaches. Tomorrow it will need to read intent: is the person passing by, asking for help, handing something over? Earning the right to that interaction requires statistical evidence, on the order of 10,000 operating hours per behavior, gathered from real fleets under today’s constraints. No company can fund tens of thousands of robots on investor capital just to collect data. The business has to pay for the learning.

That is the flywheel: contracts fund fleets, fleets generate evidence, evidence expands the safety case, and the safety case opens the next market. Because Digit is multipurpose, each new capability unlocks new workflows, and each deployment makes the underlying AI better.

The ladder to more markets

Agility’s roadmap is deliberately incremental: bins and totes, then palletizing and depalletizing, each picking and kitting, manufacturing support, grocery back rooms, shelf stocking at night, and only then construction sites, hospitals and homes, once hundreds of thousands of deployed robots have driven costs down and capability up. The company frames the expansion as widening circles of human contact: Digit v4 interacts only with other machines, Digit v5 is designed to work alongside trained adults, and future generations will meet the general public.

None of this incrementalism should be mistaken for modest ambition. Agility’s ultimate goal is the humanoid promise itself: robots in homes, helping with elder care and the tasks people would rather not do themselves. Hurst calls it part of the dream, not an alternative to the industrial business. And that end goal is precisely what drives the whole company to excellence today. Because the destination is robots working among people, every function, from safety engineering to fleet service, must be built to a standard far beyond what a caged industrial machine would require. The humanoid ambition is not a marketing story layered onto a logistics business; it is the forcing function for excellence in safety, reliability and trust.

Nor is the goal a mechanical human. Hurst wants human-centric robots: two legs, an upright torso, two arms and a face, enough form to operate in spaces built for people and to be read by them. He imagines a future closer to Monsters, Inc. than to imitation humans: many shapes and morphologies, all skipping the uncanny valley.

A company-building challenge

Hurst’s summary of the race is deliberately unglamorous. Humanoid robotics has left the research phase; it is now a company-building challenge spanning novel hardware, software and electronics, plus manufacturing, service, support and sales. Agility has built RoboFab in Oregon, which it describes as the world’s first full-scale humanoid manufacturing facility, designed to produce up to 10,000 Digits a year with roughly 75 percent domestically sourced parts.

“The companies that are going to succeed are the ones that have use cases and customers and deploy their products. It’s not the robot that shows the absolute leading AI or does the most impressive acrobatics. That doesn’t make any money,” Hurst says.

Safety, he adds, remains the primary blocker. In the next humanoid race, the winner will not be the company with the best highlight reel. It will be the one that did its homework.