TubeReads

Inside the Startup That Powers Humanoid Robots

While Boston Dynamics robots do backflips in marketing videos, real research labs struggle to make humanoids pick up boxes reliably. Flexion, a Swiss startup, claims to have cracked the code not by building hardware, but by creating what they call «the Android for humanoids» — a universal software brain that can control any humanoid robot. They're deploying 14 different humanoid platforms in parallel, training them through simulation at scales of 4,000 virtual robots running simultaneously. But can software alone bridge the gap between impressive demos and actual deployment in warehouses and factories?

Andreas Klinger ⅹ Europe's Most Ambitious StartupsTech1 People mentioned4 Glossary terms
Video length: 18:52·Published Mar 20, 2026·Video language: English
4–5 min read·4,000 spoken words → summarized to 955 words (4x)·
Watch on YouTube ↗
1 —

Key Takeaways

1

Flexion develops universal robot control software that works across different humanoid hardware platforms, reducing setup time from years to one week per new robot model.

2

The company trains robots primarily in simulation using reinforcement learning with 4,000 parallel virtual robots, rather than relying on expensive human teleoperation demonstrations.

3

Environment interaction and fine manipulation are far harder problems than athletic movements like backflips, which explains the gap between marketing videos and deployed capabilities.

4

Flexion aims to deploy its software on robots performing useful tasks for end customers in factories or warehouses within the first half of 2024.

5

Europe risks falling behind in robotics if traditional automotive and industrial companies wait for «final proof points» before investing, as catch-up cycles could take years.

In a Nutshell

Flexion is betting that humanoid robotics will mirror the smartphone revolution: hardware will commoditize while software becomes the differentiator, and Europe's manufacturing prowess combined with AI expertise could position it to capture the software layer of a multi-billion-robot economy.

2 —

The Universal Brain for Humanoids

Flexion provides software that controls any humanoid hardware, not the robots themselves.

Flexion operates 14 different humanoid robots in their Zurich lab, but they don't manufacture a single one. Instead, they build the control software — everything from low-level motor controllers to high-level reasoning agents. When a new humanoid arrives, Flexion's software stack can get it operational within one week, compared to the years of engineering effort traditionally required for single-robot, single-task deployments.

The architecture runs on three layers. Motor control and balancing algorithms execute directly on the robot to minimize latency for micro-movements. Motion planning neural networks run in the robot's backpack computer. The highest-level reasoning layer — essentially a large language model that breaks down abstract commands like «go downstairs and open the door» — runs on external servers. This tiered approach allows robots to navigate autonomously while responding to complex instructions.

The key innovation is transferability. Once one robot learns a task in Flexion's system, that capability can be deployed across their entire fleet of different hardware platforms. This creates a compounding learning effect where each robot's training benefits all others, dramatically accelerating the development cycle compared to hardware-specific approaches.

3 —

Simulation Over Teleoperation

Training 4,000 virtual robots beats teaching through human demonstration.

IMITATION LEARNING
Human-Limited Performance
Teleoperation requires rooms full of operators wearing AR glasses to demonstrate each task hundreds of times. The robot learns to mimic these demonstrations through supervised learning. The fundamental limitation: performance can never exceed the human demonstrator's ability, and operators lack the robot's force sensing and different kinematic constraints. Quick to get basic functionality, but difficult to achieve fine-grained precision.
REINFORCEMENT LEARNING
Robot-Optimized Solutions
Flexion simulates 4,000 robots in parallel using Nvidia's Isaac Lab, running multiple orders of magnitude faster than real-time. Robots start with random actions and learn through trial and error what works, receiving rewards for progress. This approach discovers solutions optimized for the robot's actual body — movements that might look unnatural to humans but are more efficient. The robot directly feels forces and can be rewarded for optimizing interactions like door opening torque.
4 —

Current Capabilities on Display

🗺️
Autonomous Navigation
Robots build real-time 3D maps of their environment and plan paths to specified locations. An operator selects a destination on a screen; the robot autonomously finds the route, avoids obstacles, and executes the movement without direct control.
🪜
Adaptive Stair Climbing
Through sensors, humanoids detect stairs and adjust their gait automatically. They can ascend, descend, and even walk backward up stairs. The same controller works in forest environments despite never training specifically for that terrain.
💪
Dynamic Balancing
Robots maintain stability when pushed or disturbed, making constant micro-adjustments. This robustness comes from extensive simulation training across varied conditions including randomized weights, joint friction, and ground friction.
📦
Object Manipulation
Robots autonomously generate training data by repeatedly grasping objects from boxes while avoiding collisions. This self-supervised approach rapidly builds datasets for manipulation tasks without human demonstration.
5 —

The Reality Behind the Marketing

Parkour is easier than picking up boxes; interaction is the hard problem.

“Surprisingly, it's easier to make these robots do back flips or parkour or jump than doing boring manual labor reliably. To do a back flip, you don't need to look around. You don't need to interact with the environment. So, what's really hard is the environment interaction.”

— Nikita

6 —

Anatomy of a Humanoid

Battery, computer, and 29+ motors working in coordinated precision.

Total Humanoid Fleet
14 robots
Operating simultaneously in Flexion's Zurich lab across different hardware platforms
Motors Per Robot
29 actuators
Excluding hand motors: 7 per arm, 6 per leg, additional motors in waist and ankles
Parallel Simulation Scale
4,000 robots
Virtual robots training simultaneously to generate massive datasets faster than real-time
New Robot Setup Time
1 week
Time to fully configure a new humanoid model with Flexion's software versus years traditionally
7 —

Europe's Robotics Window

Manufacturing strength exists, but industrial companies must move before proof arrives.

⚠️

Europe's Robotics Window

Europe possesses precision manufacturing expertise and a massive automotive industrial base that translates well to robot production. The missing ingredient is innovative AI-based software — exactly where startups like Flexion focus. The risk: traditional automotive and industrial companies are waiting for final proof points before investing. By the time robots are reliably deployed in Chinese or American facilities, it will be too late for Europe to catch up. The window for action is now, before absolute proof arrives.

8 —

People

Nikita
Company Representative
guest
Glossary
Reinforcement learningA machine learning approach where robots learn through trial and error, receiving rewards for progress toward goals rather than mimicking human demonstrations.
TeleoperationRemote control of a robot by a human operator, often used to collect demonstration data for training through imitation learning.
Reality gapThe difference between simulated robot behavior and real-world performance; bridged through randomization of physics parameters in training.
Degrees of freedomThe number of independent movements a robot joint or mechanism can perform; more degrees allow greater versatility but increase complexity.

Disclaimer: This is an AI-generated summary of a YouTube video for educational and reference purposes. It does not constitute investment, financial, or legal advice. Always verify information with original sources before making any decisions. TubeReads is not affiliated with the content creator.