Humanoid robots are often shown performing impressive single actions. Flexion Robotics is aiming at a harder and more practical target: getting a robot to carry out a chain of ordinary office chores without a person steering every movement.
The Swiss startup, founded by former Nvidia robotics researchers, has developed a training approach for robots that can handle tasks involving basic physical abilities such as opening doors, climbing stairs and carrying boxes. The larger idea is not just to make a humanoid move well, but to help it decide which learned skill to use at each stage of a job.
Why Office Chores Are a Serious Robotics Test
Many humanoid robot demonstrations focus on a narrow assignment, such as folding shirts or loading shelves. Those examples can be useful, but they do not necessarily prove that a robot will cope well in a new setting.
The source article describes a common method behind many demonstrations: teleoperation, where a person controls the robot behind the scenes. Flexion says that approach is not reliable enough when the robot has to operate in an unfamiliar environment. Its alternative is to train robots in simulation and use limited human instruction.
That distinction matters because office work is not one repeated motion. A simple request can require navigation, manipulation, balance and sequencing. A robot may need to pass through doors, use stairs, ride an elevator, find an object, unpack it and place items in the right location.
Flexion’s demonstration is built around exactly that kind of multi-step job. A modified Unitree humanoid robot receives this command:
“A parcel with snacks has been delivered for Flexion. Retrieve it using the stairs and come up using the elevator. Then unpack it and place the items into the empty drawer on the shelf in the snack area.”
According to the source, the robot then operates autonomously. The point is not just that it can walk or pick something up, but that it can connect several capabilities into one task.
How Flexion’s System Breaks Down the Work
Flexion’s approach combines multiple AI systems. The main AI model uses videos of humans performing different activities to learn what actions should happen and when. Those videos do not teach the physical mechanics of the movement. Instead, they help the model understand the order and purpose of actions.
Once the system has chosen what should happen next, it activates skills the robot has learned in simulation. Those skills are then used in the real world. In an office setting, the system may infer that reaching a mail room requires opening certain doors and using the elevator.
The software also handles motor control. That means it controls the robot’s motors so the machine can walk, move its limbs and keep its balance. In practical terms, the system needs both high-level planning and low-level physical control.
The pieces described in the source include:
- a main AI model that decides the sequence of actions;
- video-based learning that helps the model understand what to do and when;
- simulation-trained skills for physical tasks;
- motor control for walking, limb movement and balance.
This layered setup is central to Flexion’s claim. A robot that only knows one isolated task is limited. A robot that can choose among learned skills may be more adaptable when the job changes.
Reinforcement Learning Is the Core Ingredient
Nikita Rudin, the cofounder and CEO of Flexion and a former robotics research scientist at Nvidia, says the software’s “secret ingredient” is its extensive use of reinforcement learning. The source explains reinforcement learning as a way to train computers through trial and error.
In Flexion’s system, reinforcement learning is used across the software stack. That includes the master AI model, the simulation layer and the motor control layer. The result is an architecture where each level is trained to improve through repeated attempts.
This matters because humanoid robots operate in the physical world, where small errors can quickly cascade. A robot may understand that it needs to open a door, but it still has to approach it, position its body, move its limbs and maintain balance. A workplace task becomes difficult because planning and motion have to work together.
Flexion’s demonstration suggests that the value of humanoid robots may depend less on the shape of the machine and more on the intelligence coordinating it. That is also the view of George Chowdhury, an analyst with ABI Research who follows the humanoid market.
“The humanoid itself isn’t the interesting, revolutionary thing, rather it’s the AI models that back them,”
ABI Research estimates that the market for robot foundation models could be worth $150 billion by 2036.
Why the Software Layer Could Decide the Market
Tech industry leaders like Elon Musk and Jensen Huang argue that humanoids could have a major economic impact because they may eventually replace a significant share of human labor. But the Flexion example points to a condition behind that claim: humanoids need much stronger AI to become broadly useful.
A robot body that can move is only part of the product. The harder commercial question is whether the robot can be programmed to handle useful work in changing environments. Flexion is positioning its software as a possible answer to that problem.
Rudin says Flexion is working with several robotics companies and that its software works across different humanoid forms. Because there are many systems on the market, software that can operate across different bodies could become more commercially valuable.
Chowdhury also notes that Flexion will need close relationships with hardware manufacturers and will face intense competition. His broader point is blunt: without the ability to program humanoids in the way Flexion is demonstrating, “there isn’t really a market here.”
That is the clearest implication of the demonstration. The future of humanoid robots may not be determined by whether they can dance, run or perform one controlled task on camera. It may depend on whether AI systems can turn ordinary instructions into reliable sequences of physical action in real workplaces.