We are looking for a rare hybrid profile — someone who combinesdeep learningexpertisein world foundation models, generative video, and transformerswithhands-on AV simulation experience. You understand both how to train and adapt large generative models (think Cosmos, Cosmos-Transfer, diffusion-based video models, latent world models)andhow to ground them in map data and scenario semantics so the output is actually useful for training and validating perception and planning stacks.

This is not a pure simulation role, and it is not a pure ML research role.It is the bridge between the two and that bridge is where HERE's differentiation lives.

What you will do: World Foundation Models & Generative Scenario Synthesis

• Drive the technical direction for map-grounded world foundation models: how we condition generative video and world models using map data, drive data, and scenario semantics.

• Train, fine-tune, and adapt generative models (diffusion, latent video, transformer-based world models) for driving scenario generation, including domain adaptation, controllability, and conditioning on structured inputs (maps, trajectories, agentbehaviours, weather, lighting).

• Evaluate and extendstate-of-the-artfoundation models such as NVIDIA Cosmos / Cosmos-Transfer and comparable open-source world models, assessing fit for AV training data generation.

• Own the full ML lifecycle end-to-end: data curation, model training, evaluation, iteration, and the path to production-grade pipelines.

Strategic role

• Lead proof-of-concept initiativesdemonstratingmap-grounded synthetic scenario generation with key technology partners.

• Define measurable success criteria that go beyond visual realism focusing on ML training data utility, controllability, and sim-to-real transfer.

• Deliver POC outcomes with clear GO / PIVOT / NO-GO recommendations backed by quantitative evidence.

Simulation, Scenario Generation & Sim-to-Real

• Bridge generative world models with classical simulation stacks (CARLA, NVIDIA Drive Sim,AlpaSim) where structured, physics-grounded scenarios are needed.

• Author and programmatically generateOpenSCENARIO/OpenDRIVEdefinitions that feed both classical simulators and generative pipelines.

• Drive sim-to-real strategy: measure domain gap,identifyfailure modes, and define acceptable thresholds for downstream model training.

Quality Frameworks for Synthetic Training Data

• Define what "good enough" synthetic data means for AVperceptionand planning: when is photorealismrequired, when is label consistency sufficient, when does controllability matter most?

• Establish validation frameworks combining objective metrics (distribution coverage, label accuracy, FID-style measures, downstream task performance) with expert evaluation protocols.

• Specify sensor fidelity requirements: noise models, lens distortion, lidar return characteristics and how generative models should or should not reproduce them.

Technical Collaboration

• Interface with ML research teams on generative model architecture, controllability, and conditioning strategies.

• Collaborate withperceptionand planning teams to ensure synthetic data measurably improves real-world model performance.

• Translate business requirements into technical feasibility assessments for product and executive stakeholders.

Who are you?

This role requires depth inbothdeep learning and AV simulation.We are not looking for a pure simulation engineer, and we are not looking for a generalist ML researcher without AV grounding.

Must-Have: Deep Learning & Generative Models

• Proven experience training deep learning models end-to-end, with clear ownership across data, training, evaluation, and iteration.

• Expertise ingenerative video, world models, or related generative AI research/engineering.

• Deep working knowledge of diffusion models, latent video models, and/or transformer-based world models.

• Experience with high-dimensional temporal orspatio-temporal data (video, multi-sensor fusion, driving data).

• Strong Python andPyTorchengineering fundamentals; comfortable building research-grade tooling that can scale toward production.

• Demonstrated ability to take ML models from research into production, navigating real-world constraints, quality, and safety requirements.

Must-Have: AV Simulation & Scenario Domain

• 5+ years combined experience spanning AV simulation, perception/ML for AVs, or robotics simulation with meaningful exposure to both simulation platforms and ML model development.

• Hands-on experience with at least one major simulation platform: CARLA, NVIDIA Drive Sim, or equivalent.

• Fluency withOpenDRIVEandOpenSCENARIO: can author and generate scenario definitions programmatically andunderstandsmap format specifications.

• Understanding of AV testing workflows: scenario-based validation, ASAM OpenX standards, and awareness of frameworks such as ISO 34502.

• Understanding of what scenarios stress-test AVperceptionand planning systems, and why.

Must-Have: Synthetic Data Quality & Sim-to-Real

• Ability to evaluate synthetic data for ML training utility: distribution diversity, label consistency, edge-case coverage, downstream task performance.

• Experience withsynthetic-to-real transfer, domain adaptation, or closing the sim-to-real gap in a measurable way.

• Clear point of view on trade-offs between photorealism, label accuracy, controllability, and computational efficiency.

Nice-to-Have

• Hands-on experience with NVIDIA Cosmos, Cosmos-Transfer, or comparable world foundation models.

• Reinforcement learning experience, particularly where it measurably improved real-world performance.

• Experience with end-to-end driving models.

• Automotive, OEM, or other safety-critical ML deployment experience (ISO 26262, SOTIF awareness).

• Strong publicationrecordin generative models, world models, or AV ML; or significant contributions to open-source ML tooling.

• Game engine experience (Unreal, Unity) for rendering and sensor simulation pipelines.

• Experience withPyTorchLightning or similar large-scale training infrastructure.

Personal Attributes

• Bridge-builder:fluent translator between ML researchers, simulation engineers, AV domain experts, and product managers.

• Hands-on:youvalidateassumptions by training models and running simulations, not by writing specs.

• Quality-obsessed:you define objective standards where others see subjective judgments.

• Pragmatic:you balance "state-of-the-art realism" against "measurably useful for training."

• Systems thinker:you understand how every choice in data generation propagates into downstream model performance.