Three humanoid funding stories landed on the same Tuesday, and each one embodies a completely different theory of how this industry gets built. In New York, Agility Robotics hosted a livestreamed Analyst and Investor Day ahead of its Q4 public listing, armed with more than $300 million in multi-year Digit 5 orders. In Seoul, Doosan Robotics was formally seated as lead on two government-backed physical AI programs worth roughly $74 million. And in San Francisco, a company nobody had heard of before this morning announced approximately $10 million in pre-seed funding led by General Catalyst to build humanoid robots for “the world’s most dangerous jobs.”
By 2026’s inflated humanoid standards, $10 million is a rounding error. Figure sits near a $39 billion valuation, Apptronik has banked more than $935 million across its Series A rounds, and Business Insider reported last week that FieldAI is set to quintuple to a $10 billion valuation in just over a year on the strength of its robot foundation models alone. Minerva Humanoids is betting roughly one-thousandth of Figure’s valuation that the industry’s most contested assumption, that humanoids must chase labor substitution at massive scale, is not the only path to a real business. Its robot, named Roger, is aimed at offshore oil and gas operations, high-voltage electrical infrastructure, explosive ordnance disposal, and hazmat response: jobs where the alternative to a machine is not a cheaper human, but a dead one.
The thesis deserves a serious look, because if it works, it inverts almost every piece of received wisdom about humanoid unit economics.
The smallest round of the day has the sharpest thesis
Minerva was founded at the end of 2025 by CEO Sandor Felber, who previously helped launch MIT CSAIL’s humanoid robotics research line and worked on Tesla’s Optimus program, and CTO Maurice Rahme, who spent six years at Boston Dynamics as a technical lead on Stretch, taking it from prototype to commercially deployed product. That founding profile, one Optimus alumnus and one Boston Dynamics shipping veteran, is the standard pedigree now, but the product framing is not.
Roger is explicitly semi-autonomous. A trained specialist operates it through a VR headset, seeing through the robot’s cameras and commanding fine manipulation directly, while Minerva Intelligence, the company’s onboard AI layer, handles balance, fall recovery, and navigation. The company says operators learn full-body control in minutes, and it frames the human as remaining “in command of every critical action.”
The round was led by General Catalyst with Long Journey Ventures and Credo Ventures as co-leads, plus a long tail of participants that includes Hugging Face, an unusual name in a hardware pre-seed and a signal of where the talent and capital networks around physical AI are converging. The company went from idea to a walking humanoid in roughly five months, with prototypes manufactured in Sindelfingen, Germany, and plans to produce in both Germany and the United States “using a Western industrial base without reliance on single-source Chinese components.” It has already demonstrated at the NATO Explosive Ordnance Disposal Centre of Excellence in Slovakia, established research partnerships with MBZUAI for energy-sector humanoids and the University of Waterloo for perception in smoke, darkness, and sensor-denied conditions, and completed on-site testing at two of the world’s ten busiest airports. First paid pilots begin this fall.
Why hazardous work inverts the humanoid math
The dominant humanoid business model, pursued with variations by Figure, Agility, Apptronik, and Tesla, is labor substitution in logistics and manufacturing. The pitch is arithmetic: a robot-hour replaces a wage-hour, so the business only works when autonomy is good enough to keep the human-to-robot supervision ratio low and the fleet utilization high. Every point of autonomy short of that target bleeds margin, which is why warehouse humanoids are autonomy-first by financial necessity.
Hazardous work runs the math backwards. Minerva’s pitch is not that Roger replaces an offshore technician or an EOD operator; it is that Roger lets that specialist do the same job from a command post. “We didn’t set out to replace the people doing them,” Felber said in the announcement. “We set out to build a robot that lets them do the job efficiently from the comfort of their office or home, without asking them to risk their lives for our safety or our energy.”
The cost baseline is also different. Minerva estimates, by applying the ILO’s global fatal work accident rate to the IEA’s 76 million-person global energy workforce, that an energy worker dies on the job roughly every 75 minutes, about 7,000 deaths a year. The figure is the company’s own extrapolation rather than a direct count, but the methodology is disclosed and defensible, and the direction is not in dispute: rigs, substations, and hazmat scenes kill and maim workers at rates warehousing does not approach. Against that baseline, a robot that fails on 30 percent of task attempts may still be a strict improvement, because the fallback is not a scanner gun, it is a person in a protective suit walking toward a device. Dominic Asquith, a British Special Forces veteran who now leads hazardous operations at Minerva, put it with the bluntness of the trade: “Every EOD operator knows the long walk. When the robot can’t complete the task and the operator has to put on the suit and approach the device alone.”
That asymmetry does three things to the business model. It relaxes the autonomy requirement, because a human is in the loop on critical actions rather than excluded from them. It tolerates one operator per robot, because the value created is risk absorption, not wage arbitrage: you are not paying the robot to be cheaper than the specialist, you are paying it to be destroyable in ways the specialist is not. And it changes the buyer. The relevant comparators are not hourly wages but incident costs, insurance exposure, evacuation downtime on offshore platforms, and the political cost of a fatality, none of which are cheap. NATO units and energy majors are not price-sensitive in the way a 3PL is.
Five months from blank page to walking machine
The other claim worth interrogating is speed. Building a functional biped in five months, bespoke design, manufacturing, and assembly included, would have been impossible five years ago. Today it is merely very hard, and the explanation is that the humanoid component ecosystem has matured into something close to a commodity layer: actuators, harmonic and planetary roller drivetrains, tactile sensors, vision modules, and compute boards are all purchasable, by the hundred, from specialist suppliers in Germany, Japan, and increasingly the United States.
That maturity cuts both ways. Apptronik CEO Jeff Cardenas warned publicly last week that actuators can represent up to 60 percent of a humanoid’s bill of materials and that America still lacks the parts base to build these machines at scale domestically. The same supply chain that lets two founders ship a biped in five months also concentrates cost and dependency in exactly the components that are hardest to source, a squeeze the memory market is now applying to brains as well. Minerva’s answer is positioning: Western sovereign, German and American production, no single-source Chinese components. For defense and critical-infrastructure buyers, that sentence is not marketing, it is a procurement requirement, and it is the strongest structural moat a $10M company can plausibly claim.
The day’s other two bets, and what they reveal
Minerva’s announcement lands in deliberate contrast with the two larger stories of the same day. Agility’s Analyst and Investor Day is a public-market ritual: presentations from CEO Peggy Johnson, Schaeffler’s humanoid robotics president, a Foxconn strategic investment director, and AUVSI’s acting chief executive, all building the case for the Churchill Capital Corp XI combination expected to close in Q4, after which Agility trades on Nasdaq as AGLT. The metrics that matter to that audience are deployment-hours (65,000 and counting), contracted revenue (over $300 million in multi-year Digit 5 orders), and unit economics. “Humanoid robotics is moving from what’s possible to what’s practical,” Johnson said, which is precisely the sentence a company about to be marked daily by public investors must be able to say.
Doosan’s programs represent the state-capital model: 98.9 billion won across two projects, 68.1 billion won of it government-funded, spanning on-device AI cobots, an industrial humanoid with two-arm manipulation and omnidirectional mobility, and imitation-learned welding for nuclear power equipment with a 95 percent path-similarity target, commercialization targeted for 2031. That is patient, consortium-built, certification-first development with an explicit national capability objective, and it does not need to satisfy a venture return timeline.
Public equity demands deployment metrics now. State capital accepts 2031. And venture pre-seed capital, Minerva’s lane, buys concentrated theses with fast falsification: one robot, one segment, paid pilots within months of stealth.
What could kill this thesis
The bear case is straightforward. Ten million dollars is thin for hardware iteration; robots that operate on offshore platforms need corrosion resistance, ingress protection, and potentially explosion-proof certification, each a multi-quarter engineering program on its own. VR teleoperation over satellite links from an offshore rig imposes latency that no “low-latency control loop” press language fully dissolves. The incumbent competition, decades of tracked single-arm EOD robots, is cheap, battle-proven, and already fielded by thousands of units, and Minerva’s claim that two-handed dexterity changes the mission set is exactly that, a claim, until independent operators validate it. And the addressable market is structurally bounded: the world has thousands of EOD teams and offshore installations, not millions of sites, so even success looks like a high-price, low-volume business rather than the fleet-scale story warehouse players tell.
The honest assessment is that Minerva has not yet proven the thesis; it has proven only that the thesis is now testable at pre-seed cost. The pilots this fall at energy sites and the two airports under exploration are the experiment.
What to watch
Three checkpoints will tell us whether risk transfer is a real category or a fundable narrative. First, Minerva’s paid pilots this fall: whether specialists actually choose Roger over the tracked incumbents, and whether the two-handed dexterity advantage survives contact with real devices. Second, Agility’s public filings as AGLT in Q4, which will give the industry its first public-grade unit economics for a deployed humanoid fleet, a benchmark every private player’s deck will quietly reprice against. Third, whether the NATO EOD community, which has watched single-arm robots plateau for two decades, converts its showcase interest into procurement.
The humanoid industry has spent three years arguing about whose robot will replace the most warehouse workers. The $10M round suggests a sharper question: who gets the jobs where replacement was never the point, and survival was.