About Flexion

At Flexion, we're building the intelligence layer powering the next generation of humanoid robots. Our mission is to accelerate the transition from fragile prototypes to real-world humanoid deployment. We are founded by leading scientists in robot reinforcement learning (ex-Nvidia, ex-ETH Zürich), and backed by leading international VC firms. In just months, we’ve gone from our first line of code to deploying real humanoid capabilities.

The role

We are seeking an expert in dexterous manipulation and large-scale modeling to lead the development of our physical foundation models. The goal of this position is to leverage internet-scale egocentric video to build Vision-Language-Action (VLA) models that enable our humanoid robots to interact with the world with human-like fluidity. You bring a deep understanding of how to bridge the gap between observing human actions in video and executing high-DOF (20+) motor control.

As a Research Engineer for Dexterous Manipulation at Flexion, you’ll work in our Zürich office to develop and deploy state-of-the-art learning-based controllers. You will take ownership of the model architecture, integrating egocentric priors with real-time robot policies, to ensure our hardware can manipulate objects reliably and flexibly in unstructured environments.

Key Responsibilities

• Scalable Egocentric Pre-training: Architect and implement large-scale pre-training objectives for egocentric video datasets to learn generalizable representations of hand-object interactions and spatial-temporal dynamics.

• VLA Foundation Modeling: Develop and scale multi-modal Foundation Models that unify visual perception and natural language instructions into actionable robotic trajectories.

• Generative Policy Design: Design and optimize generative action heads using Diffusion Models and Flow-matching techniques to capture the multi-modal distribution of complex human movements.

• Humanoid Motion Alignment: Develop novel algorithms to align human-centric video representations with the kinematic constraints of 20+ DoF humanoid systems, ensuring fluid and stable execution.

• Reinforcement Learning & Fine-tuning: Utilize Offline RL and high-fidelity simulation fine-tuning to optimize foundation model performance for high-success-rate physical manipulation.

• Cross-Functional Research: Translate cutting-edge research in scaling laws and world models into production-ready architectures that enhance robot reliability and autonomy.