Swiss Startup Trains Humanoid Robots Using Layered AI Skills

A Swiss robotics startup founded by former Nvidia researchers has developed a new approach to training humanoid robots that could address one of the industry's biggest challenges: getting machines to perform multi-step tasks in unfamiliar environments without constant human supervision.

Flexion Robotics trains its robots by teaching them individual skills in simulation, then deploying a master AI algorithm that determines when and how to apply those skills in real-world scenarios. The company demonstrated its system with a modified Unitree humanoid that successfully completed a complex office errand: retrieving a delivered package by navigating stairs and an elevator, then unpacking items into a designated storage area—all from a single natural language command.

A departure from teleoperation

Most current humanoid demonstrations rely on teleoperation, where human operators control the robot's movements in real time. This method works for specific, rehearsed tasks but fails when robots encounter unexpected situations or need to chain together multiple actions independently.

Flexion's system operates differently. The main AI model analyzes videos of humans performing various activities, then matches those observations to skills the robot has already learned in simulation. When tasked with reaching a mailroom, for instance, the model recognizes it needs to open doors and operate an elevator—skills it practiced virtually before attempting them physically.

Reinforcement learning at every layer

According to Nikita Rudin, Flexion's cofounder and CEO, the software's distinguishing feature is its extensive use of reinforcement learning throughout the entire stack. This training method, which teaches systems through trial and error, powers everything from the high-level planning model down to the motor control systems that manage walking, balance, and limb movement.

The layered architecture allows the robot to build competence incrementally. Rather than learning each complete task from scratch, it develops a library of fundamental capabilities that can be recombined for new situations.

Why it matters

The humanoid robotics market hinges on software more than hardware. ABI Research estimates the market for robot foundation models could reach $150 billion by 2036, but analyst George Chowdhury notes that without effective programming approaches like Flexion's, "there isn't really a market here." Companies deploying humanoids need systems that can adapt to varied workplace environments without extensive reprogramming for each new task—exactly the capability that has remained elusive with traditional training methods.

Commercial strategy and competition

Flexion is working with multiple robotics manufacturers and designed its software to function across different humanoid platforms. This hardware-agnostic approach could prove commercially valuable given the growing number of humanoid designs entering the market.

However, Chowdhury cautions that success will require close collaboration with hardware makers and that Flexion faces significant competition in a crowded field. The company's ability to deliver on its simulation-based training promise will determine whether it can capture a meaningful share of the emerging market.

These details were first reported by WIRED.