About the job
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 Research Engineer to own the robotics vertical of our world models: taking our video-native foundation models and turning them into policies that control real robots in the real world.
Responsibilities
Design and execute end-to-end robot learning pipelines — from task design and demonstration data collection through policy training, and physical evaluation
Deploy and iterate on learned policies (VLAs, diffusion policies, World Action Models) on real robot hardware, closing the loop between model predictions and physical outcomes
Run controlled experiments to understand how world model representations, data composition, and fine-tuning strategies translate to downstream manipulation and locomotion performance
Build and maintain physical evaluation benchmarks and infrastructure — designing tasks, procuring hardware, calibrating systems, and measuring real-world success rates
Coordinate robot data collection efforts across internal teams and external partners, ensuring data quality, coverage, and consistency across embodiments
Partner with the world model research team to translate model capabilities into concrete robotics applications, identifying where our video foundation models unlock new robot behaviors
Identify and resolve bottlenecks across the robotics stack — whether in data, training infrastructure, hardware configuration, or evaluation methodology — to keep the overall system moving fast
Qualifications
Minimum
Hands-on robotics experience spanning data collection, model training, and physical evaluation. Direct experience with modern learned policies (e.g., VLAs, diffusion policies) on real hardware.
Experience with robot data collection, teleoperation, and demonstration pipelines across at least one manipulation or mobile platform
Strong intuition for the full robot learning lifecycle: task design → data collection → policy training → physical evaluation
Comfort working across software, hardware, and physical systems — you can debug a training run and reconfigure a robot workspace in the same afternoon
Proficiency with at least one ML framework (e.g., PyTorch, JAX)
Preferred
Bonus: experience with video or multimodal generative models, world models, or using foundation model representations for downstream control