About the role
*Open to candidates based near our NYC office or those willing to relocate.
Building general world models — systems that understand and simulate reality across tasks, modalities, and domains — requires closing the loop between learned representations and real-world action. We're looking for a Robotics Engineer to bring our video-native foundation models onto real hardware: deploying world-model-based policies on real robots and making them work in the real world.
You will work across the full stack of robot learning — from data collection and task design, to policy deployment, to physical evaluation. This is a hands-on, execution-oriented role at the intersection of foundation models, learned robot policies, and hardware. You'll bring deep robotics domain expertise and help us ship world-model-based robot policies end-to-end, with applications ranging from manipulation to mobile robotics.
What you'll do
- Own the deployment loop for learned policies (VLAs, diffusion policies, World Action Models) on real robot arms: inference serving, action decoding, controller integration, latency budgets, and safety limits
- Diagnose real-world policy failures — determining whether the problem is data coverage, camera/proprioception mismatch, or execution — and partner with the research team to fix it at the right layer
- Design demonstration data collection protocols (task setup, camera placement, teleop conventions) that produce data our world models can actually learn from — working with our existing teleop and lab infrastructure
- Partner with research to run controlled experiments connecting world model representations, data composition, and fine-tuning recipes to physical success rates
- Define and run physical evaluation protocols so results are trustworthy and comparable across policy iterations
- Adapt our policy recipe to partner hardware and new embodiments
What you'll need
- You have personally deployed software on real robot hardware and iterated to make it work — whether learned policies, motion planning, classical control, or perception-driven manipulation. You can walk through a specific system you shipped, what broke on the physical robot that didn't break in sim or testing, and how you fixed it.
- Fluency in the software-to-robot interface: action/command spaces, control frequencies, observation pipelines, calibration, and how each affects real-world behavior
- Comfort working at the intersection of software and physical systems: you can reconfigure a robot workspace and trace a policy failure in the same afternoon
- Bonus: direct experience deploying learned policies (VLAs, diffusion policies) on real hardware; experience with video/multimodal generative models or world models