Research Engineer – Embodied AI

Role overview This position focuses on advancing how foundation models are trained and adapted for physical, real-world tasks. The work sits at the boundary between research and engineering: studying promising techniques, forming testable hypotheses, implementing them rapidly, and running disciplined experiments to verify whether they yield meaningful improvements in robot capability. The emphasis is on measurable physical performance rather than novelty for its own sake.

Responsibilities - Investigate new methods for policy learning, model adaptation, and embodied intelligence on physical systems. - Develop and evaluate techniques for post-training vision-language-action and other multimodal foundation models. - Explore architectural choices, training objectives, data mixtures, representations, and learning algorithms for physical tasks. - Apply and compare imitation learning, reinforcement learning, offline learning, and behavior cloning approaches. - Design controlled experiments that isolate the factors driving performance gains and interpret ambiguous results. - Reimplement published research papers from scratch and validate that results reproduce before extending them. - Build evaluation methodologies that surface genuine capability differences between models, separating signal from noise. - Analyze policy failures, form hypotheses about root causes, and translate findings into the next round of experiments. - Collaborate with simulation, data, and systems engineers to convert research ideas into complete training pipelines. - Track and synthesize relevant progress in embodied AI, multimodal learning, and foundation models.

Requirements - Strong grounding in machine learning, deep learning, optimization, and experimental methodology. - Solid software engineering skills, with the ability to ship clean research code. - Hands-on experience training models in PyTorch, JAX, or comparable frameworks. - Ability to read a research paper, extract the core idea, and implement it independently. - Intuition for experimental design and an ability to judge whether an apparent improvement is real or an artifact. - Comfort working on open-ended problems where the solution path is undefined.

Nice to have - Prior work in embodied AI, robot learning, or sequential decision-making. - Research experience with vision-language-action models, multimodal models, or large foundation models. - Familiarity with reinforcement learning, imitation learning, or offline RL. - Background in generative modeling, world models, or representation learning. - Experience training large models or running distributed training jobs. - Publications or substantive open-source contributions in relevant machine learning areas. - Experience taking a research prototype toward real-world deployment on physical hardware.