Research Engineer, Robot Learning

Role overview

This role focuses on training and improving large-scale robot learning systems that connect multimodal observations with physical action. The work spans the full model lifecycle, from preparing datasets and developing training methods through evaluation and deployment on real robots. It combines hands-on research engineering with systematic experimentation, failure analysis, and production-minded training infrastructure.

Responsibilities

- Train robot-learning and Vision-Language-Action models using multimodal data such as demonstrations, trajectories, video, language, robot state, and actions. - Own pre-training, post-training, fine-tuning, and adaptation workflows for robot policies and foundation models. - Develop and evaluate data mixtures, sampling methods, filtering, augmentation, curricula, and objectives for imitation learning, behavior cloning, reinforcement learning, and offline learning. - Implement policy architectures, run large-scale experiments and ablations, and analyze training dynamics, regressions, and behavioral failure modes. - Improve generalization across tasks, objects, environments, and robot embodiments while optimizing distributed training for stability, throughput, and GPU utilization. - Build reproducible experiment infrastructure and carry models through evaluation to deployment on physical robots in collaboration with robotics, simulation, data, and evaluation specialists.

Requirements

- Strong software engineering and machine-learning fundamentals, including the ability to implement research ideas and evaluate them quickly. - Experience with robot learning, multimodal foundation models, transformers, representation learning, or related areas. - Ability to diagnose whether a model limitation is caused by data, optimization, architecture, or training objectives. - Practical familiarity with one or more of imitation learning, reinforcement learning, offline RL, Vision-Language-Action models, diffusion or flow-based policies, world models, or distributed training. - Comfort working across experimentation, data pipelines, evaluation, and deployment rather than focusing on only one stage of model development.

Nice to have

- Experience training robot foundation models or large models through pre-training and post-training. - Experience with large robot-trajectory datasets, manipulation policies, multi-robot or cross-embodiment data, or physical-robot deployment. - Experience with JAX, large-scale PyTorch systems, distributed GPU training, or strong publications and open-source work in robot learning, embodied AI, or machine learning.

What success looks like

Training runs are reproducible and scalable, new data and training recipes produce measurable capability gains, and policies generalize to new tasks, objects, environments, and embodiments. Training failures become diagnosable, and improvements observed during experiments translate into better behavior on real robots.