Machine Learning Engineer, Synthetic Data
Salary not listed • London, United Kingdom • Full-time
Posted 2d ago
About the job
We are seeking a Machine Learning Engineer to join our Simulation Teams, focusing on advancing end-to-end autonomous driving research. This role is crucial for accelerating our AV2.0 development by incubating next-generation world model architectures and the synthetic data they produce. You will be responsible for building, scaling, and optimizing these models, integrating them into high-throughput generation and training infrastructure. Your work will directly impact the speed at which we can train, evaluate, and deploy driving models, bridging the gap between ML research and engineering by collaborating with scientists on architecture and platform engineers on infrastructure.
Responsibilities
- Post-train, distill, and evaluate GAIA-family large video diffusion models to improve generation quality, controllability, and generalization.
- Run ablation studies and investigate failures in visual quality, temporal consistency, motion, and action conditioning.
- Build and run synthetic video generation pipelines across tens of thousands of driving segments.
- Optimize inference efficiency, manage GPU resources, and ensure data quality throughout the pipeline.
- Integrate synthetic data across different stages of driving-model training, develop sampling strategies, and measure its impact.
- Collaborate with driving-model teams to determine effective data mixtures and with simulation and evaluation teams to turn generated scenarios into useful closed-loop tests.
- Translate advances in world modelling, video generation, flow matching, distillation, and data-centric training into measurable improvements.
- Post-train and iterate GAIA-class world models for synthetic-data capabilities, including embodiment transfer and pose transfer.
- Own the generation loop from configuration to training-ready artifacts, ensuring clear lineage.
- Integrate synthetic data into driving-model training (behavior cloning, reward models, RL), including binarization, mix ratios, and quality filters.
- Diagnose and fix geometry, artifacts, calibration, and controllability failures in generated video.
- Improve throughput and yield through inference optimizations and self-serve workflows.
- Expand coverage to new vehicle platforms and safety-critical scenarios.
- Partner with world-model researchers, infra, and driving-model owners to ensure generation, evaluation, and training remain a unified system.
Requirements
- 4+ years in applied ML / research engineering with a track record of training and shipping neural nets.
- Strong Python and PyTorch (or equivalent) skills.
- Comfort with GPU training, debugging, and reading model code.
- Hands-on experience with video, generative, or world models (e.g., diffusion, flow-matching, autoregressive video, novel-view synthesis, neural rendering).
- Working knowledge of cameras and 3D geometry (multi-camera rigs, intrinsics/extrinsics, warps/reprojection).
- Evidence of taking generated or simulated data into a trained downstream model and measuring impact.
- Ability to operate generation or training at real scale (multi-GPU jobs, workflow orchestration, large video artifacts).
- Collaborative and experimental working style with researchers and platform engineers.