About this role.
Nebius AI R&D is seeking a Senior or Staff ML Engineer to join their Physical AI research team, focusing on building intelligent agents that can perceive, reason, and act in the physical world. The role involves modifying large foundation models and learning algorithms for robotic agents, prototyping in simulation, and validating on real-world systems. Key research areas include vision-language-action models, reinforcement learning, imitation learning, world models, and sim-to-real transfer. This is a research-heavy position requiring deep expertise in ML, robotics, and large-scale model training, with opportunities to publish and collaborate across teams. The ideal candidate is a hands-on engineer excited about bridging cutting-edge AI research with practical robotic systems.
Role DNA
A quick view of the complexity, pace, ownership and collaboration implied by the job description.
Job Complexity
5/5Pace & Pressure
4/5Autonomy Level
4/5Communication Load
4/5Salary analysis
Estimated compensation compared with the broader US market for similar roles.
Core skills
Skills and capabilities most closely associated with this opportunity.
Cover letter sample
Dear Hiring Manager,
I am excited to apply for the Senior ML Engineer (AI Research, Physical AI) position at Nebius. With a strong background in machine learning, reinforcement learning, and robotics, I have spent the last several years developing and deploying large-scale models for perception and control. My experience includes building vision-language-action models and implementing sim-to-real pipelines, which directly align with the core research areas of your team.
In my previous role, I led a project that integrated multimodal transformers with robotic manipulation policies, achieving a 40% improvement in task success on real-world dexterous tasks. I am passionate about bridging the gap between simulation and reality, and I have extensive experience with distributed training frameworks like PyTorch and DeepSpeed. I am particularly drawn to Nebius because of its unique position at the intersection of cloud infrastructure and applied AI research, offering an opportunity to scale my work from prototype to production.
I thrive in collaborative research environments and have a track record of publishing at top conferences (e.g., CoRL, ICRA) while also shipping production code. I am eager to contribute to your Physical AI research and help build intelligent agents that can perceive, reason, and act in the physical world. Thank you for your consideration.
Sincerely, [Your Name]
Sample interview questions
In my previous role, I worked on a VLA model that processed camera images and language instructions to output motor commands. One major challenge was aligning visual features with action spaces, as the model needed to understand spatial relationships and generate precise control signals. We addressed this by using a cross-modal attention mechanism and incorporating proprioceptive feedback. Another challenge was data efficiency; we used a combination of human teleoperation data and simulation to pretrain the model, then fine-tuned it on a small set of real-world demonstrations. We also implemented robust data augmentation to handle domain shifts.
First, I would identify the critical factors that affect real-world performance, such as object dynamics, friction, sensor noise, and lighting. I would then create a simulation environment that models these factors with high fidelity, using tools like MuJoCo or Isaac Gym. To bridge the sim-to-real gap, I would use domain randomization—randomizing textures, forces, and camera angles—to make the policy robust to variations. I would also implement a systematic evaluation protocol that tests the policy in both simulation and on the real robot, iterating on the simulation parameters based on real-world failures.
I once had to train a deep RL agent for a manipulation task that required millions of environment interactions. I used a distributed setup with multiple workers collecting experience in parallel and sending data to a central replay buffer. For the policy update, I used a GPU cluster with synchronous gradient updates, leveraging PyTorch's distributed data parallel. I also used Ray RLlib to manage the worker processes and resource allocation. To speed up experience collection, I parallelized the environment execution across CPU cores. This setup reduced training time from days to hours, allowing faster experimentation and iteration.
Imitation learning is great for leveraging human demonstrations and bootstrapping a policy, especially when the reward function is hard to define. It's efficient for tasks with clear trajectories. However, it can suffer from distribution shift and lacks the ability to improve beyond the demonstrated behavior. RL, on the other hand, allows exploration and can find better policies, but it can be sample-inefficient and requires careful reward design. In practice, I often combine both: pretrain a policy using behavioral cloning from demonstrations, then fine-tune with RL to optimize for a specific objective. The choice depends on the task, data availability, and the ability to design a reward function. For exploration-heavy tasks with sparse rewards, I'd lean on RL with reward shaping, while for tasks where demonstrations are readily available and the dynamics are complex, I'd start with imitation learning.
I regularly read papers from top conferences like CoRL, ICRA, NeurIPS, and RSS, and I follow researchers and labs in the field. I also actively participate in open-source projects and experiment with new methods on my own to validate their feasibility. When deciding which ideas to pursue, I consider three factors: the potential impact on the field, the feasibility with our existing resources, and the alignment with our team's long-term research goals. I believe in focusing on problems that are both scientifically interesting and practically relevant, and I often use a small-scale proof-of-concept to test the viability of a new idea before committing significant resources.
About Nebius:
Nebius is leading a new era in cloud infrastructure for the global AI economy. We are building a full-stack AI cloud platform that supports developers and enterprises from data and model training through to production deployment, without the cost and complexity of building large in-house AI/ML infrastructure.
Built by engineers, for engineers. From large-scale GPU orchestration to inference optimization, we own the hard problems across compute, storage, networking and applied AI.
Listed on Nasdaq (NBIS) and headquartered in Amsterdam, we have a global footprint with R&D hubs across Europe, the UK, North America and Israel. Our team of 1,500+ includes hundreds of engineers with deep expertise across hardware, software and AI R&D.
The role
This role is for Nebius AI R&D, a team focused on applied research in AI. Our Physical AI research aims to build intelligent agents that can perceive, reason, and act in the physical world. Research areas include:
Vision-language-action models for general-purpose robotic control
Reinforcement and imitation learning from human demonstrations, simulation, and real-world experience
Scalable collection, generation, and curation of multimodal embodied data
Simulation, world models, and sim-to-real transfer
Multimodal sensing, including vision, touch, force, and proprioception
You will modify large foundation models and learning algorithms for robotic agents, prototype new capabilities in simulation, and validate promising approaches on real-world systems. The results will often lead to collaboration with adjacent research, infrastructure, and engineering teams, where findings are scaled and applied in practice.
We are currently looking for senior- and staff-level ML engineers to work on research in areas such as:
Vision-language-action models and multimodal foundation models for robotics
Reinforcement learning, imitation learning, and learning from demonstrations
Scalable acquisition and generation of human, robot, and simulated interaction data
World models, planning, and model-based control
Sim-to-real transfer, domain adaptation, and robust policy evaluation
Dexterous manipulation, whole-body control, and general-purpose robotic agents
Some examples of what your responsibilities might include are:
Designing, implementing, training, and evaluating large models and learning algorithms for robotic agents
Developing vision-language-action architectures that connect multimodal perception and language understanding with physical control
Investigating reinforcement learning and imitation learning methods for sparse, delayed, or difficult-to-verify objectives
Building scalable methods for incorporating demonstrations, teleoperation data, video, simulation trajectories, and autonomous robot experience into foundation models
Designing capture methodologies, datasets, evaluation protocols, and data-quality pipelines for embodied learning
Developing simulation environments and conducting sim-to-real experiments on physical robotic platforms
Exploring planning, guided generation, and search over action trajectories
Prototyping new capabilities in areas such as dexterous manipulation, mobile manipulation, and whole-body control
Writing robust research software and distributed training infrastructure that enable rapid experimentation
Collaborating with research and engineering teams to translate promising ideas into reliable real-world systems
Communicating results through technical reports, open-source releases, demonstrations, and research publications
We expect you to have:
A profound understanding of the theoretical foundations of machine learning, reinforcement learning, or robot learning
Deep expertise in at least one relevant area, such as reinforcement learning, imitation learning, multimodal generative modeling, computer vision, robotics, planning, or control
Experience training and evaluating modern deep learning models, including transformer-based or multimodal foundation models
Substantial experience training large models across multiple computational nodes
Strong software engineering and algorithm-design skills; we primarily use Python
Deep experience with a modern deep learning framework; we primarily use JAX
Experience designing, executing, and analyzing machine learning experiments with appropriate statistical rigor
Ability to formulate meaningful research questions, design experiments that test clear hypotheses, and draw defensible conclusions
Experience implementing research ideas and iterating quickly across modeling, data, infrastructure, and evaluation
Strong communication and leadership abilities, including the ability to collaborate across research and engineering disciplines
Ability to document research findings clearly and contribute to technical reports or research publications
Nice to have:
Experience working with real-world robots and robotic simulation environments
Experience with dexterous manipulation, whole-arm manipulation, mobile manipulation, or humanoid robotics
Experience with multimodal sensing, including tactile, force-torque, depth, and proprioceptive signals
Experience collecting human demonstrations through teleoperation, motion capture, wearable devices, or observation
Experience developing or post-training vision-language models, vision-language-action models, or video and world models
Experience with deep reinforcement learning techniques such as offline RL, actor-critic methods, PPO, reward modeling, preference learning, or model-based RL
Familiarity with robotics tools and simulators such as MuJoCo, Isaac Sim, Isaac Lab, PyBullet, ROS, or equivalent systems
Knowledge of scalable training techniques such as FSDP or ZeRO, FlashAttention, mixed-precision training, quantization, and distributed checkpointing
A PhD in Computer Science, Robotics, Machine Learning, Artificial Intelligence, or a related technical field, or equivalent practical experience
A track record of impactful publications, open-source contributions, or deployed robotic systems
Experience engineering large distributed data-processing, simulation, or model-training systems
A record of building and delivering products or research prototypes in a dynamic, startup-like environment
Passion for moving research from controlled experiments to capable, reliable real-world robotic systems
Excellent command of English, with strong technical writing, presentation, and communication skills
Proficiency in contemporary software engineering practices, including version control, testing, code review, and CI/CD
Benefits & Perks:
- Competitive compensation
- Career growth and learning opportunities
- Flexibility and ownership
- Collaborative and innovative culture
- Opportunity to work on impactful AI projects
- International environment and talented teams
What’s it like to work at Nebius:
Fast moving – Bold thinking – Constant growth – Meaningful impact – Trust and real ownership – Opportunity to shape the future of AI
Equal Opportunity Statement:
Nebius is an equal opportunity employer. We are committed to fostering an inclusive and diverse workplace and to providing equal employment opportunities in all aspects of employment. We do not discriminate on the basis of race, color, religion, sex (including pregnancy), national origin, ancestry, age, disability, genetic information, marital status, veteran status, sexual orientation, gender identity or expression, or any other characteristic protected by applicable law.
Applicants must be authorized to work in the country in which they apply and will be required to provide proof of employment eligibility as a condition of hire.
If you need accommodations during the application process, please let us know.
Annual salary information is not provided for this position. Explore salary ranges for similar roles in our Salary Directory ›
Upload your resume now
- To unlock remote work opportunities and be discovered by global employers.
This job listing has been manually reviewed by the Jobicy Trust & Safety Team for compliance with our posting guidelines, including verification of the company's legitimacy, accuracy of job details, clarity of remote work policy, and absence of misleading or fraudulent content.
