Senior AI/ML Research Engineer – Model development

Job description

Primary Function of Position We are building advanced augmented dexterity for next-generation robotic platforms. As a Senior AI/ML Research Engineer, you will develop and fine-tune the foundation models—VFMs, VLMs, and VLA models—that let our Embodied-AI system understand the surgical scene and act within it. Within a hierarchical, multimodal stack, you will own the model layer: adapting large pretrained vision and multimodal models on surgical data to extract anatomy, instruments, actions, and context from intraoperative video, and connecting perception to reasoning and action.

Partnering with the broader AI/ML team, you will drive the path from offline research to robust, real-time performance in the OR. Working within Intuitive's Future Forward research organization, you will identify, build, and fine-tune the AI/ML models and algorithms that let us deliver safe and performant embodied-AI systems. This role calls for someone equally comfortable getting hands-on with models and data and designing systems that scale.

Roles and Responsibilities Develop, fine-tune, and evaluate the AI/ML models—including foundation and multimodal models—that enable the system to perceive the surgical scene and translate intent and observations into safe, performant behavior. Establish strong baselines by reproducing relevant state-of-the-art approaches, then iteratively advance them with in-house models and components while keeping interfaces stable.

Build and maintain training and data pipelines that combine real demonstration data with simulation, and establish human-in-the-loop pipelines for continuous model improvement. Define and run evaluation for model performance, repeatability, and safe failure/abort behavior, and establish the path from offline evaluation on recorded data to robust, real-time integration. Partner with data and annotation teams to shape label taxonomies, quality control, and the data pipeline that feeds the models.

Collaborate across AI/ML research, robotics, software, and data engineering to align on interfaces and deliver models that enable rapid prototyping and learning while building toward a product solution.

Qualifications

Minimum Qualifications

MS or PhD in CS, EE, Robotics, or a related field, with 5+ years of applied AI/ML research experience in areas such as robot learning, embodied AI, control, or sequential decision-making. Hands-on experience training policies or models from data, including imitation/behavior cloning and reinforcement learning, and fine-tuning pretrained models. Experience with vision-action (VA), vision-language-action (VLA), or goal/intent-conditioned models, including models that produce action or control outputs.

Familiarity with world models and self-supervised predictive architectures (e.g., JEPA-style models, MAE, DINO) for learning dynamics and latent representations to support planning and control. Comfort building training and evaluation loops over both simulated and real-world data. Strong software and ML-engineering skills in Python and C++, with proficiency in one or more of PyTorch/TensorFlow/JAX.

A research-and-prototyping mindset

comfortable working in ambiguity, framing open-ended problems, running rapid experiments, and reading and reproducing recent papers to pull promising techniques into practice. Sound judgment about the path from prototype to product: writing code others can build on, knowing when to optimize versus when to move fast, and thinking ahead about data quality, evaluation, and robustness even at the research stage. Solid foundations in linear algebra, probability, and optimization, enough to reason about and debug model behavior from first principles.

Comfort collaborating across a multidisciplinary team (ML, robotics, software, and clinical/domain experts) and communicating tradeoffs and findings clearly.

Preferred Qualifications

Modern policy architectures: diffusion policies, transformer policies, action chunking (e.g., ACT), and generalist robot policies (RT-X / OpenVLA-style). DAgger / human-in-the-loop and data-flywheel pipeline experience. Sim-to-real transfer and domain randomization (e.g., NVIDIA Isaac Sim). Teleoperation, kinematics, and real-time on-robot deployment. Publications at CoRL, RSS, ICRA, IROS, or NeurIPS.