Note: The job is a remote job and is open to candidates in USA. HERE Technologies is a leading company at the intersection of mapping and AI technology, seeking an AV Simulation Domain Expert. The role involves bridging deep learning with AV simulation to enhance the creation of synthetic driving scenarios, ensuring effective training and validation of perception and planning systems.
Responsibilities
- 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, agent behaviours, weather, lighting)
- Evaluate and extend state-of-the-art foundation 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
- Lead proof-of-concept initiatives demonstrating map-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
- Bridge generative world models with classical simulation stacks (CARLA, NVIDIA Drive Sim, AlpaSim) where structured, physics-grounded scenarios are needed
- Author and programmatically generate OpenSCENARIO / OpenDRIVE definitions that feed both classical simulators and generative pipelines
- Drive sim-to-real strategy: measure domain gap, identify failure modes, and define acceptable thresholds for downstream model training
- Define what 'good enough' synthetic data means for AV perception and planning: when is photorealism required, 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
- Interface with ML research teams on generative model architecture, controllability, and conditioning strategies
- Collaborate with perception and planning teams to ensure synthetic data measurably improves real-world model performance
- Translate business requirements into technical feasibility assessments for product and executive stakeholders
Skills
- Proven experience training deep learning models end-to-end, with clear ownership across data, training, evaluation, and iteration
- Expertise in generative 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 or spatio-temporal data (video, multi-sensor fusion, driving data)
- Strong Python and PyTorch engineering 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
- 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 with OpenDRIVE and OpenSCENARIO: can author and generate scenario definitions programmatically and understands map 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 AV perception and planning systems, and why
- Ability to evaluate synthetic data for ML training utility: distribution diversity, label consistency, edge-case coverage, downstream task performance
- Experience with synthetic-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
- 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 publication record in 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 with PyTorch Lightning or similar large-scale training infrastructure
Company Overview
Company H1B Sponsorship