The posting
HEAD OF PHYSICAL AI
Location: Remote — Europe and the United States preferred; exceptional candidates globally will be considered Employment: Full-time Reports to: CEO Role type: Hands-on technical leader and team builder Travel: As needed
ABOUT OUR CLIENT
Our client builds the data, evaluation, and deployment layer for Physical AI.
The company works across multimodal robot and human data, annotation and assurance, model evaluation, and the systems that turn physical-world experience into useful robot behavior.
Miraxis is hardware- and model-agnostic. What matters is whether a dataset, model, or method produces a measurable improvement on a real task. The company will build focused model and evaluation capabilities where they strengthen its data products, demonstrate the value of its data, or solve a clear customer or partner problem.
THE ROLE
Our client is looking for a Head of Physical AI to establish and lead its AI research and engineering function.
You will decide which Physical AI problems the company pursues, define how results are evaluated, and remain directly involved in the most important technical work. You will connect four areas that are often treated separately:
- Multimodal and embodied data
- Transformer-based models and robot policies
- Rigorous offline and real-world evaluation
- Deployment on physical systems
This is a player-coach role. During your first year, at least half of your time will be spent on direct technical work: designing models and experiments, writing or reviewing code, inspecting data, debugging training runs, analyzing failures, and reviewing robot rollouts.
You will also build a small, focused team of researchers and engineers as the work requires it.
WHAT YOU’LL DO
- Define a focused Physical AI research and engineering roadmap with clear hypotheses, baselines, milestones, success measures, and stop criteria.
- Select the model families Miraxis should train, adapt, or evaluate and determine when to build internally, use open models, license technology, or work through partners.
- Personally design, adapt, train, and evaluate Transformer-based systems for embodied tasks.
- Work across areas such as vision-language-action models, multimodal Transformers, robot foundation models, action representation, imitation learning, reinforcement learning, world models, cross-embodiment transfer, and robot-policy evaluation.
- Define the sensors, modalities, annotations, data mixtures, coverage, and quality controls required to train and evaluate selected models.
- Measure how data quality, diversity, and composition affect model behavior and real-world task performance.
- Establish reproducible offline and real-world evaluation systems, baselines, held-out conditions, and release gates.
- Protect evaluations against leakage, overfitting, and weak or misleading success criteria.
- Take projects from problem definition through training, hardware integration, and real-world validation.
- Analyze failures across data, perception, models, control, hardware, and the operating environment.
- Design safe, staged physical testing and deployment plans with clear supervision and rollback mechanisms.
- Recruit and lead a small team of complementary researchers and engineers.
- Lead architecture, experiment, code, rollout, and failure reviews.
- Translate customer and partner needs into testable technical requirements.
- Communicate technical strategy, evidence, uncertainty, and limitations clearly to customers, partners, and investors.
WHAT WE’RE LOOKING FOR
DEEP, HANDS-ON TRANSFORMER EXPERTISE
This is a hard requirement. You must have personally made material architecture or training decisions in at least one substantial Transformer-based system, such as:
- Vision Transformers
- Vision-language or vision-language-action models
- Multimodal foundation models
- Decision Transformers
- Diffusion Transformers
- Video Transformers
- World models
- Transformer-based perception, planning, or control systems
You should be able to explain how you represented and tokenized inputs and outputs, fused modalities, structured attention and temporal context, selected losses, built data mixtures, distributed and monitored training, diagnosed failures, and changed the system to improve task performance.
Using hosted model APIs, prompting language models, or running an unchanged public training recipe does not meet this requirement.
PHYSICAL AI EXPERIENCE
You have worked on machine learning for a system that perceives or acts in the physical world, such as robotics, autonomous vehicles, drones, industrial automation, manipulation, mobile robots, humanoids, or wearable and egocentric systems.
At least one substantial project must have progressed beyond offline datasets or simulation into a real or operational physical system. Simulation experience qualifies only when paired with credible sim-to-real ownership and physical validation.
HANDS-ON TECHNICAL ABILITY
- Strong Python engineering skills
- Direct experience with PyTorch or an equivalent deep-learning framework
- Ability to read and debug unfamiliar model and training code
- Experience designing controlled experiments and analyzing rollout failures
- Experience working with large multimodal datasets
- Sound judgment around compute, memory, training stability, inference latency, and cost
TECHNICAL LEADERSHIP
You have set the direction for a significant research, model-development, robotics, or cross-functional technical program. You have made architecture and resource decisions, mentored or hired technical talent, stopped weak lines of work, and helped take a result into deployment.
A management title is not required, but direct technical ownership is.
STRONGLY PREFERRED
Candidates combining deep Transformer expertise with strong computer-vision experience will receive priority. Relevant areas include:
- Visual representation learning
- VLMs and video or temporal modeling
- Detection, segmentation, and tracking
- Multi-view and egocentric vision
- 3D and spatial reasoning
- Pose estimation, calibration, and localization
- Sensor fusion
- Real-time vision systems
Additional valuable experience includes:
- Vision-language-action models or robot foundation models
- Action tokenization or continuous action generation
- Diffusion or flow-matching policies
- Imitation learning or reinforcement learning
- World models and cross-embodiment training
- Robot manipulation
- Distributed training and inference optimization
- ROS 2
- Sim-to-real transfer
- Safety-critical systems
- Widely used open-source work or personally owned research publications
- Early-stage company experience
- Work with technical customers or research partners
A PhD in machine learning, computer vision, robotics, computer science, or a related field is valuable but not required. An equivalent record of original model work, technical leadership, and real-system delivery is equally relevant.
Miraxis cares most about what you personally designed, trained, evaluated, and deployed.
WHAT SUCCESS LOOKS LIKE
Within 90 days:
- Audit Miraxis’s data, evaluation assets, partnerships, and model opportunities.
- Select one or two focused research bets.
- Establish a reproducible model baseline and initial evaluation suite.
- Define a credible path to physical validation.
- Present a practical 12-month roadmap supported by working technical evidence.
Within six months:
- Train, adapt, or rigorously evaluate at least one relevant Transformer-based model or robot policy.
- Confirm or reject at least one important technical hypothesis.
- Establish a repeatable data-to-training-to-evaluation workflow.
- Test on a real robot, directly or through a credible hardware partner.
- Document representative failures and their implications.
Within twelve months:
- Demonstrate a measurable improvement attributable to Miraxis data, methods, or assurance systems.
- Close the loop between model failures, data decisions, retraining, and re-evaluation.
- Deliver a model, benchmark, evaluation system, or deployment that a customer or partner can use.
- Build a small team capable of running the work without unnecessary process layers.
Success will be judged by decision quality, reproducibility, real-system results, and customer value—not by team size, paper count, parameter count, or experiment volume.
