Key Notes
Synthetic-data ownership spans GAIA-class world models, large-scale GPU generation and the training stack that consumes generated driving experience. Relevant expertise includes 4+ years in applied ML or research engineering, strong Python and PyTorch, hands-on video or world-model work, camera and 3D-geometry knowledge, downstream-model evaluation and multi-GPU workflow experience. The London role is full time under Wayve’s hybrid policy.
What You'll Work On
- Post-train GAIA-class world models for rig transfer, pose transfer, dashcam restaging and other synthetic-driving capabilities.
- Run the generation pipeline from configuration through large-scale GPU inference to traceable, training-ready synthetic artefacts.
- Integrate synthetic experience into behaviour-cloning, reward-model and reinforcement-learning training with controlled mix ratios and quality filters.
- Diagnose geometry, calibration and controllability failures across camera intrinsics, extrinsics, warps, odometry and video artefacts.
- Improve inference throughput, valid-generation yield and self-service workflows while expanding synthetic coverage to new vehicle platforms and safety-critical scenarios.
Why This Role Matters
- Wayve’s driving-model training stack uses synthetic experience alongside real driving data, so generation quality and lineage determine whether new vehicle embodiments and behaviours can be incorporated into training credibly.
- New vehicle platforms and safety-critical scenarios can be explored before large real-world datasets exist, making synthetic-data throughput and quality material to how quickly those cases enter model development.
What They Are Looking For
- Experience: Relevant experience includes 4+ years in applied ML or research engineering, with a record of training and shipping neural networks.
- Technical: Strong Python and PyTorch or equivalent, including GPU training, debugging and model-code work, plus reliable multi-GPU generation or training at real scale, are sought.
- Technical: Hands-on experience with video, generative or world models such as diffusion, flow matching, autoregressive video, novel-view synthesis or neural rendering is sought.
- Domain: Working knowledge of cameras and 3D geometry, including multi-camera rigs, intrinsics, extrinsics and reprojection, is sought.
- Experience: Evidence of taking generated or simulated data into a trained downstream model and measuring its impact is sought.
- Preferred: Controllable generation, inference-speed optimisation, AV or robotics simulation, production research workflows, reward models, cloud GPU fleets and distributed training are desirable.