Founding Research Scientist - World Models and Self-Play

Mission

We are developing a physics-based world model to serve as a universal cognitive layer for versatile robot fleets independent of hardware or embodiment. Our core hypothesis is that a world model trained through self-play incorporating physics priors and recalibrating with real robotic rollouts outperforms methods limited to human demonstration data. You will lead significant research efforts determining the success of this concept.

Responsibilities

• Lead and design research on architectures for world models including dynamics models, physics-informed priors, and self-play curricula targeted at robotic control.
• Develop and operate large-scale self-play training loops, analyze their failures, and continuously improve robustness.
• Manage comprehensive scaling experiments and ablation studies from initial hypothesis formation through final documentation and publication.
• Work hands-on with rollout data from our robot fleets (Unitree G1/B2W, ROSbot 3, Z1) to bridge the simulation-to-reality gap effectively.
• Collaborate with academic partners such as TUM and MIRMI, publish findings where appropriate, and contribute to shaping the research direction as an early team member.

Requirements

• Extensive research experience in reinforcement learning, world modeling, or model-based control, evidenced by a PhD or equivalent industrial research background.
• Practical expertise working with self-play methodologies, model-based reinforcement learning, or learned dynamics models. Backgrounds in video-generation or physics-informed learning are also valued.
• Ability to independently tackle open-ended research challenges with minimal supervision in a nascent and dynamic team environment.
• Capability to understand and analyze robotics or physics simulation code, even if not the core focus.

Preferred Qualifications

• Authored leading research papers on world models, self-play, model-based RL, or generative modeling in video/3D domains.
• Experience with SE(3)-equivariant neural architectures or other geometric and structured priors.
• Track record mentoring junior researchers or informally leading small research teams.
• Previous practical experience in performing sim-to-real transfers on physical robotic hardware.