Flexion Robotics raises $50M from DST Global and Nvidia to build simulation-trained intelligence layer for robots
Nov 20, 2025 with Nikita Rudin
Key Points
- Flexion Robotics raises $50 million from DST Global, Nvidia, and others to build an intelligence layer that abstracts visual inputs and delegates common-sense reasoning to large language models, letting robots train on simulated data at scale.
- The Zurich startup sidesteps the sim-to-real fidelity problem by having robots learn from visual abstractions rather than raw camera feeds, so simulation need not be photorealistic to transfer skills to the physical world.
- Flexion expects physics-based simulators to remain the primary training environment longer than the industry assumes, with generative AI accelerating asset creation rather than replacing simulation itself.
Summary
Flexion, a Zurich-based robotics startup, has raised $50 million in its first disclosed funding round, with DST Global and Nvidia participating alongside Verci, Promus, and Moonfire.
The company's CEO and co-founder describes Flexion as building the intelligence layer that runs across robot types, from humanoids to mobile manipulators. The core technical bet is that simulation, not physical data collection, will supply the majority of training data for robot policies. Flexion uses reinforcement learning to run robots through millions of simulated attempts, generating what it describes as tens or hundreds of years of simulated experience, then transfers those learned behaviors to the real world.
The sim-to-real gap
The standard criticism of simulation-trained robots is fidelity: you can't perfectly model every blade of grass or door texture, so policies trained in simulation break down in deployment. Flexion's answer is to avoid training directly on raw camera inputs. Instead, it uses models like Meta's Segment Anything to abstract the visual environment, effectively normalizing what a robot "sees" so that all doors, regardless of texture or lighting, look functionally the same. That abstraction means the simulation doesn't need photorealistic accuracy to produce transferable skills.
On the common-sense layer, the approach is to not build it at all — Flexion delegates that to existing large language models. Asked what's trash on a table versus what should stay, GPT-class models already handle that reliably. Flexion treats that as a solved problem and focuses its training effort on the physical manipulation layer.
Looking forward, the CEO expects physics-based simulators like Unreal Engine to remain the primary training environment for longer than the industry assumes, with generative AI accelerating asset creation — producing the varied objects, terrains, and environments robots need to encounter — rather than replacing the simulators themselves. Gaussian splats offer a parallel path, capturing real-world environments quickly to build simulation asset libraries.
Flexion is based in Zurich — its outdoor robot footage was shot in the Swiss Alps — and the CEO is in San Francisco scouting for a second office.