What appears to be a Unitree humanoid robot equipped with Flexion’s AI systems. | Source: Flexion
Flexion Robotics AG last week said it has raised Series A funding of $50 million. The company is building a reinforcement learning and sim-to-real platform that can power humanoid robots across morphologies and tasks.
Over the past few years, generative AI has changed how many people code, analyze data, and reason. At the same time, developers have looked for ways to apply this same power to robotics, noted Flexion. With the flexibility of new artificial intelligence models, roboticists could break free from the brittle, task-specific systems that rely on scripted behaviors, said the Zurich-based company.
Flexion is using generative AI and large language models (LLMs) to build models that can automate tasks involving reasoning, writing, and creativity. Founded in 2024, the startup said its full autonomy stack spans:
- Command layer: Language models for common-sense reasoning take the tasks described in natural language, break them down into subtasks, and provide the necessary environment understanding and grounding.
- Motion layer: This includes a vision-language-action (VLA) model trained primarily on synthetic data, fine-tuned for real-world edge cases.
- Control layer: Transformer-based, low-latency whole-body control with a modular skill library enables rapid composition of new behaviors.
Flexion claimed that its approach allows robots to be deployed with minimal human involvement.
Flexion Reflect v0 takes a step toward generalizable autonomy
Flexion said its AI architecture starts with LLM and vision language model (VLM) agents for task scheduling and common sense. These agents decompose goals, select tools, and understand everyday conventions. Users can program desirable outcomes through prompting and fine-tuning.
Next, it uses a general motion generator. Given images, 3D perception, and an LLM-produced task instruction, the system can propose short-horizon and collision-aware local trajectories. For example, it can tell the robot to move an end effector to grasp a box or for its full body to navigate.
Finally, it uses a reinforcement learning (RL)-based whole-body tracker. The system can execute commands across all terrain types and different command spaces.
Flexion asserted that this modularity avoids “end-to-end monoliths” and improves generalization by keeping interfaces clean and testable. In addition, its data strategy is asymmetric. The company said it uses simulation and synthetic data wherever possible, and it selectively incorporates real data to close specific gaps.
Startup raises second round this year
Flexion’s Series A round included participation from DST Global Partners, NVentures (NVIDIA‘s venture capital arm), redalpine, Prosus Ventures, and Moonfire. It followed $7.35 million in seed funding from Frst, Moonfire, and redalpine just a few months earlier.
Flexion said that it plans to use the latest financing to expand its Zurich research and development team, scale compute and robot fleets, establish a U.S. presence, and accelerate the commercialization of its autonomy stack.
The company is already working with major OEM partners, and it said the funding will help scale these partnerships globally.
Tell Us What You Think!