Last week, Flexion Robotics AG announced it has secured $50 million in Series A funding. The company is developing a reinforcement learning and sim-to-real platform designed to enable humanoid robots to handle a wide range of forms and tasks.
Bringing Generative AI to Robotics
In recent years, generative AI has transformed how people code, analyze data, and solve problems. Flexion noted that developers are now exploring ways to bring this same capability to robotics. With the adaptability of modern AI models, roboticists can move beyond fragile, task-specific systems that depend on pre-programmed behaviors, the Zurich-based company said.
Flexion leverages generative AI and large language models (LLMs) to create systems capable of automating tasks that require reasoning, writing, and creative thinking. Founded in 2024, the startup stated that its complete autonomy stack includes:
Command layer: Language models handle common-sense reasoning by interpreting tasks described in natural language, decomposing them into subtasks, and providing the required environmental understanding and context.
Motion layer: A vision-language-action (VLA) model, initially trained on synthetic data and fine-tuned for real-world edge cases, powers this layer.
Control layer: Transformer-based, low-latency whole-body control, combined with a modular skill library, allows for the quick creation and composition of new behaviors.
Flexion stated that its method enables robots to operate with little to no human intervention.
Flexion Reflect v0 Advances Toward General Autonomous Capabilitie
Flexion explained that its AI architecture begins with LLM and vision-language model (VLM) agents for task planning and common-sense reasoning. These agents break down goals, choose appropriate tools, and interpret everyday norms. Users can shape desired outcomes through prompts and fine-tuning.
The next layer is a general motion generator. By combining images, 3D perception, and LLM instructions, it generates short-horizon, collision-aware trajectories for tasks like grasping or full-body navigation.
Finally, a reinforcement learning (RL)-based whole-body tracker executes commands across diverse terrains and control domains.
Flexion emphasized that this modular design avoids “end-to-end monoliths” and enhances generalization by keeping interfaces clean and testable. Its data strategy is asymmetric: simulations are prioritized, with real-world data added only to fill gaps.
Startup Secures its Second Funding Round of the Year
Flexion’s Series A round saw participation from DST Global Partners, NVentures (NVIDIA’s venture arm), redalpine, Prosus Ventures, and Moonfire, following a $7.35 million seed round from Frst, Moonfire, and redalpine just a few months earlier.
The company said it will use the new funding to grow its Zurich R&D team, expand compute resources and robot fleets, establish a U.S. presence, and accelerate the commercialization of its autonomy stack.
Flexion, already collaborating with major OEM partners, noted that the funding will support scaling these partnerships globally.
Read the original article on: The Robot Report
Read more: Using Magnetic Fields, Soft Robots Can now Operate More Intelligently On Their Own