What seems to be a Unitree humanoid robotic geared up with Flexion’s AI methods. | Supply: Flexion
Flexion Robotics AG final week stated it has raised Sequence A funding of $50 million. The corporate is constructing a reinforcement studying and sim-to-real platform that may energy humanoid robots throughout morphologies and duties.
Over the previous few years, generative AI has modified how many individuals code, analyze knowledge, and cause. On the similar time, builders have seemed for tactics to use this similar energy to robotics, famous Flexion. With the flexibleness of recent synthetic intelligence fashions, roboticists may break away from the brittle, task-specific methods that depend on scripted behaviors, stated the Zurich-based firm.
Flexion is utilizing generative AI and huge language fashions (LLMs) to construct fashions that may automate duties involving reasoning, writing, and creativity. Based in 2024, the startup stated its full autonomy stack spans:
- Command layer: Language fashions for common sense reasoning take the duties described in pure language, break them down into subtasks, and supply the required atmosphere understanding and grounding.
- Movement layer: This features a vision-language-action (VLA) mannequin skilled totally on artificial knowledge, fine-tuned for real-world edge instances.
- Management layer: Transformer-based, low-latency whole-body management with a modular ability library allows speedy composition of recent behaviors.
Flexion claimed that its strategy permits robots to be deployed with minimal human involvement.
Flexion Replicate v0 takes a step towards generalizable autonomy
Flexion stated its AI structure begins with LLM and imaginative and prescient language mannequin (VLM) brokers for job scheduling and customary sense. These brokers decompose objectives, choose instruments, and perceive on a regular basis conventions. Customers can program fascinating outcomes by way of prompting and fine-tuning.
Subsequent, it makes use of a basic movement generator. Given pictures, 3D notion, and an LLM-produced job instruction, the system can suggest short-horizon and collision-aware native trajectories. For instance, it may inform the robotic to maneuver an finish effector to know a field or for its full physique to navigate.
Lastly, it makes use of a reinforcement studying (RL)-based whole-body tracker. The system can execute instructions throughout all terrain varieties and completely different command areas.
Flexion asserted that this modularity avoids “end-to-end monoliths” and improves generalization by retaining interfaces clear and testable. As well as, its knowledge technique is uneven. The corporate stated it makes use of simulation and artificial knowledge wherever attainable, and it selectively incorporates actual knowledge to shut particular gaps.
Startup raises second spherical this 12 months
Flexion’s Sequence A spherical included participation from DST International Companions, NVentures (NVIDIA‘s enterprise capital arm), redalpine, Prosus Ventures, and Moonfire. It adopted $7.35 million in seed funding from Frst, Moonfire, and redalpine only a few months earlier.
Flexion stated that it plans to make use of the newest financing to broaden its Zurich analysis and improvement staff, scale compute and robotic fleets, set up a U.S. presence, and speed up the commercialization of its autonomy stack.
The corporate is already working with main OEM companions, and it stated the funding will assist scale these partnerships globally.
