Flexion Robotics teaches robots new skills without humans, using only AI and simulation

Swiss startup Flexion Robotics has developed a method for teaching humanoid robots new tasks without human involvement. Instead of human demonstrators and data annotators, it uses AI and virtual simulation.

Routine matters more than stunts

Today's humanoid robots can put on an impressive show: they run marathon distances, perform on stage, and show off acrobatic moves. Videos of these stunts rack up millions of views and fuel the image of "robots of the future" already living among us.

Behind the flashy demos, though, lies a more mundane reality. The real value of these machines comes down to their ability to handle monotonous, physically demanding, or dangerous tasks — the kind that tire people out, injure them, or that they simply refuse to do. Assembling components on a production line, packing and sorting goods, moving heavy loads, servicing warehouse racks: that is what industry needs humanoids for.

Teaching a robot each specific task is still expensive and slow. The traditional cycle involves human demonstrators, data annotators, and verification engineers, and the whole process takes months. Flexion Robotics set out to remove humans from that chain.

Simulation instead of an instructor

The startup's core idea is to move the entire training cycle into a virtual environment. AI algorithms generate thousands of ways to perform a task on their own, score the quality of each attempt, and iteratively adjust the model's behavior. No live trainers or demonstrators are required — just a simulator and a reinforcement learning algorithm.

The approach offers several advantages:

• scaling without hiring and training a team of annotators
• safe training on dangerous, rare, and non-standard scenarios
• learning several skills in parallel
• cutting the cycle from task definition to a working skill several times over
• lowering the cost of developing each new skill

Once training in the simulator is complete, the system transfers the learned strategies to a physical robot, a process known as sim-to-real transfer. Historically, this is where the most serious problems have arisen: behavior in simulation often diverged from reality because of inaccuracies in the physics model and differences in surface friction and material response. Flexion Robotics calls the accuracy and stability of this transfer its key competitive advantage.

Competitors and context

Synthetic robot training is one of the hottest topics in the industry right now. Google DeepMind, Physical Intelligence, Figure AI, 1X Technologies, and dozens of other startups are pursuing research in the same direction. Most of them rely either on large human-generated datasets or on costly interaction with real-world environments.

Flexion Robotics is betting that a fully automated pipeline will let robots be deployed where it previously made no economic sense: small production lines, regional logistics, food processing, and agriculture. If the method proves effective in real-world conditions, the economics of robotics could change fundamentally.

What it means

Taking humans out of the training loop would make robots quick to adapt to changing tasks and widely accessible. If Flexion Robotics achieves stable sim-to-real transfer in practice, companies will be able to launch new robotic operations in weeks rather than months. Industrial automation would no longer be the preserve of large corporations with multimillion-dollar budgets.