robot learning research engineer (manipulation policies)

Mundane is a venture-backed seed-stage robot learning startup founded by a team of Stanford researchers and builders. We’re deploying a massive fleet of humanoid robots to perform mundane tasks in commercial environments, collecting data to build the next generation of embodied intelligence. We are a fast-paced, execution-driven team of engineers, roboticists, and builders. Our robots operate in real customer environments — and improve through real-world experience. About the Role You will develop and ship learning-based manipulation policies that run on real robots. Our robots already collect real-world data and execute manipulation tasks. Your role is to turn that data into policies that improve reliability, generalize across tasks, and hold up under real-world distribution shift. This role is deeply hands-on and execution-focused. You will implement models, run controlled experiments, and validate improvements directly on physical robots. Success is measured by real-world performance, not benchmark metrics. At Mundane, your models will not live in simulation or papers — they will deploy to humanoid robots operating in customer environments. What You’ll Own Development and improvement of real-world manipulation policies Policy architecture and training recipes for real-world manipulation Robustness improvements (recovery behaviors, partial observability, drift, edge cases) Experiment discipline and clear ablation methodology Scaling from single-task policies to multitask robot capabilities Responsibilities Extend and improve our policy learning stack (imitation learning / sequence-based policies) for real-world manipulation tasks Design and run disciplined experiments to improve policy performance, including clear ablations and controlled comparisons Develop multitask policies with effective task conditioning and thoughtful data mixture strategies Improve robustness through techniques such as data augmentation, recovery behaviors, and training under partial observability Design and run systematic stress tests to evaluate distribution shift, drift, and edge-case failures Work closely with infrastructure engineers to scale training pipelines and experiment workflows Collaborate with reliability engineers to define evaluation gates and deployment criteria Package trained models for deployment, addressing latency, stability, and safety constraints Investigate real-world failures and iterate rapidly to improve policy robustness Qualifications Strong PyTorch and ML engineering skills with the ability to implement and ship reliable training pipelines Practical experience with imitation learning or behavior cloning Experience training sequence-based models such as transformers, diffusion policies, or related architectures Comfort running real-world experiments and debugging issues across data, training, and deployment Strong experimental rigor, including designing ablations, maintaining reproducibility, and avoiding “demo-only” improvements Nice to Have Experience with robotic manipulation systems and real-world robot experimentation Familiarity with common failure modes in manipulation tasks Experience scaling training across large datasets or multi-GPU environments Background in embodied AI or robot learning systems What You’ll Get Direct ownership over the policies that control robots operating in real environments Early equity with meaningful upside in a venture-backed robotics company The opportunity to see your research deployed quickly on real humanoid robots Close collaboration with hardware, infrastructure, and deployment teams A front-row seat in scaling a technically ambitious robotics company from seed stage #J-18808-Ljbffr mundane