Flexion Robotics, a Swiss startup founded by former Nvidia robotics researchers, has built a software stack that lets humanoid robots carry out multi-step office chores from a single natural-language prompt. In a demonstration video, a modified Unitree humanoid running Flexion's system retrieved a delivered parcel, took the stairs down and the elevator back up, unpacked the contents, and placed the items into a drawer on a shelf in the office snack area. ABI Research estimates the market for the kind of robot foundation models Flexion is building will be worth $150 billion by 2036.
The pitch is that nearly every humanoid demo circulating today depends on teleoperation — a hidden human operator puppeteering the robot through a narrow, rehearsed task like folding a shirt or loading a shelf. That works for marketing reels and falls apart the moment the robot encounters an unfamiliar room, an unexpected door, or a chore it was not specifically scripted to perform. Flexion's argument is that no economically useful humanoid will ship until that gap closes.
Its solution is to layer several AI systems on top of each other. A master planning model digests videos of humans performing everyday tasks and works out the sequence of sub-tasks required to complete a goal. The planner then maps each sub-task to a library of individual skills — opening a door, riding an elevator, picking up a box — that the robot has separately learned inside simulation. A lower-level controller handles balance, gait, and limb movement so the robot can actually execute the plan in a real building.
Key facts
- 01Flexion Robotics, a Swiss startup founded by former Nvidia robotics researchers, trains humanoids in simulation rather than via teleoperation.
- 02ABI Research projects the market for robot foundation models will reach $150 billion by 2036.
- 03A modified Unitree humanoid running Flexion's stack autonomously fetched a parcel, navigated stairs and an elevator, and stocked a shelf from a single natural-language prompt.
- 04Every layer of Flexion's stack — the master planner, the simulation, and the low-level motor control — is trained with reinforcement learning.
- 05Flexion is hardware-agnostic and is working with multiple humanoid manufacturers rather than building its own robot.
Critically, none of those layers depend on a human in the loop at runtime. The robot decides which skill to deploy and how to chain them, which is what makes a prompt as open-ended as the snack-delivery instruction tractable.
Flexion cofounder and CEO Nikita Rudin, previously a robotics research scientist at Nvidia, says the system leans heavily on reinforcement learning at every layer of the stack — the planner, the simulation environment, and the motor controller all trained through trial and error rather than scripted by hand. That contrasts with the imitation-learning approach common in teleoperation-based demos, which tends to overfit to the specific environment and props used during data collection.
Elon Musk and Jensen Huang have both argued for years that humanoids will reshape labor markets at the scale of the personal computer or the smartphone. The bottleneck has consistently been software rather than hardware. Unitree, Figure, 1X, Apptronik, and Boston Dynamics all have capable bipedal platforms; what none of them have shipped is a generalist control layer that lets an off-the-shelf humanoid walk into an unfamiliar building and do useful work without a script.
“The humanoid itself isn't the interesting, revolutionary thing, rather it's the AI models that back them”— George Chowdhury, ABI Research analyst
George Chowdhury, an analyst at ABI Research who covers the humanoid market, frames Flexion's bet directly: the robot is the commodity, the model is the product. His firm's $150 billion 2036 forecast for robot foundation models reflects that view — value accrues to whoever owns the planner, not whoever stamps out the chassis.
Flexion is positioning accordingly. Rudin says the company is hardware-agnostic and is already collaborating with several robotics companies, with the software designed to work across different humanoid form factors. Given how fragmented the humanoid hardware market is becoming — dozens of platforms from Chinese, US, and European manufacturers — a control stack that runs on any of them is more commercially defensible than one welded to a single robot.
“there isn't really a market here”— George Chowdhury, ABI Research analyst
Chowdhury cautions that Flexion will have to integrate deeply with hardware makers to make the approach work in production, and that the competitive field is crowded with well-funded incumbents pursuing variations on the same idea. Tesla's Optimus team, Figure's Helix model, and Nvidia's own GR00T humanoid platform are all chasing the same generalist-control prize, often with far larger compute budgets. Without the ability to program humanoids in the way Flexion demonstrates, he argues, the underlying hardware business does not exist.
The demo also leaves open the questions that always trail humanoid videos: how often the system fails, how long the parcel-fetch run actually took, and how the model behaves when the elevator is occupied or the drawer is locked. Flexion has not published benchmark numbers, success rates, or pricing, and the snack-area errand is still a constrained office environment rather than a chaotic warehouse floor or hospital ward. Those gaps will close or they will not, and the second-order question — whether reinforcement learning at every layer generalizes better than imitation learning — is exactly the kind of empirical fight that will play out over the next two product cycles.
The interesting structural point is that the humanoid market is following the same pattern as the LLM market: hardware proliferates, and value concentrates in whoever owns the model on top. If ABI's $150 billion 2036 figure is even directionally right, a Swiss team of ex-Nvidia researchers shipping a hardware-agnostic planner has picked the right layer of the stack to compete on. Whether Flexion or one of the better-capitalized incumbents wins it, the takeaway for buyers is that the robot you purchase in 2030 will matter far less than the subscription you pay to make it do anything useful.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
