Waymo Machine Learning Intern at Waymo works on generative world models and controllable simulation for autonomous driving. The role includes driving architectural roadmap and training large-scale foundation models for 4D world modeling.
Responsibilities
Own the Architectural Roadmap: Drive the technical vision and long-term roadmap for foundational 4D world modeling, unified spatiotemporal tokenizers (3D/4D VAEs), video generation architectures, and generative sensor simulation across Waymo -- enabling zero-shot generalization across next-generation vehicle platforms, sensor rigs, and international geographies.
Lead Large-Scale Pre-Training: Architect, train, and scale foundation video and physical world models (including Sparse MoE and long-context block-sparse attention architectures) across distributed TPU/GPU clusters, establishing efficient training recipes, multimodal data curation strategies, and scaling laws.
Drive Advanced Post-Training & Alignment: Design and implement robust post-training paradigms, including Supervised Fine-Tuning (SFT), Reinforcement Learning with AI/Critic Feedback (RLAIF), model merging/souping, and 3D/semantic conditioning to guarantee strict physical realism, counterfactual pose-divergence stability, and fine-grained control over safety-critical scene elements (signage, traffic signals, VRU gestures, and adverse weather).
Integrate Multimodal & Physical Signals: Develop action-conditioned, 3D-grounded physical world models and multi-view generative pipelines that couple explicit geometric and physical priors (4D scene graphs, occupancy, vehicle kinematics, and interactive agent dynamics) with neural synthesis across Camera, LiDAR, and Radar—faithfully modeling physical sensor transport, occlusions, and causal state transitions.
Optimize for Production & Serving: Scale inference and evaluation frameworks into production simulation infrastructure, applying few-step and causal autoregressive distillation (e.g., Distribution Matching Distillation [DMD], Self-Forcing, Consistency/Flow distillation), quantization, and KV-cached streaming inference for 10 Hz high-throughput closed-loop validation and RL gyms.
Foster Research-to-Production Impact: Partner closely with Onboard Perception, Planner, and AI Foundation labs across Waymo and Alphabet (e.g., Google DeepMind) to adapt frontier world models into deployment-ready simulation environments that achieve downstream perception/policy parity for rare-event reproduction, synthetic data training, and closed-loop RL.
Mentor & Elevate Engineering: Serve as a technical authority, mentoring senior engineers and driving high engineering standards across training stability, reproducibility, and system design.
Education: PhD or MS in Computer Science, Machine Learning, Robotics, Computer Vision, or a related quantitative field (or equivalent practical research/industry tenure, including academic appointments looking to transition to industry).
Senior Technical Leadership: 8+ years of industry or relevant research lab experience (or tenure-track/associate professorship equivalent) developing, deploying, and leading complex ML architectures.
Large-Scale Pre-Training Track Record: Demonstrated experience training large-scale generative and physical world models from scratch (e.g., action-conditioned 3D/4D world models, large video diffusion models, autoregressive spatiotemporal transformers, flow matching, or vision-language-action models) on high-performance compute clusters (multi-node GPU/TPU).
Post-Training & Alignment Expertise: Hands-on experience designing post-training and student-conversion pipelines for generative models—including few-step diffusion/flow distillation (e.g., DMD, MMD, Consistency Models, Rectified Flow), autoregressive exposure-bias mitigation (e.g., Self-Forcing), SFT, and Reinforcement Learning (RLAIF / reward-model alignment)—to control and stabilize long-horizon rollouts.
Deep 3D/4D Vision Understanding: Deep technical foundation in 3D computer vision and physical world modeling—including action-conditioned scene dynamics, explicit 3D physical/geometric priors (4D occupancy, dynamic meshes, rigid/articulated kinematics), spatiotemporal tokenization (3D/4D causal VAEs), continuous ray/camera geometry (e.g., Plücker rays, cross-rig calibration), and generative 4D world representations (coupling explicit physical priors with neural generation).
Framework Proficiency: Expert-level proficiency in Python and modern distributed ML frameworks (such as JAX/Flax, XLA/Pallas, PyTorch, Megatron-LM, Pathways, or DeepSpeed) for multi-node model, sequence, and data parallelism.
Domain Pedigree: Prior experience from leading video generation labs, foundation model startups, advanced robotics institutes, or high-tier academic research groups focusing on physical AI, foundation video models, or world simulation.
Multi-Modal Sensor Simulation: Background in generative modeling and synthesis of non-RGB physical sensors—specifically 3D LiDAR point clouds/range images and 4D imaging radar—alongside multi-camera automotive rigs across diverse vehicle platforms, weather, and lighting conditions.
Embodied AI & World Dynamics: Hands-on experience in physical world modeling and embodied dynamics—including action-conditioned causal world models, differentiable or hybrid physics engines, rigid/articulated body kinematics, vehicle dynamics, interactive multi-agent traffic dynamics, or closed-loop robotics simulators.
High-Performance Infrastructure: Experience with low-level systems optimization, XLA/Pallas TPU kernels or custom CUDA/Triton kernels, block-sparse / Flash / Splash attention over 100k+ token sequences, and high-throughput KV-cached distributed serving.
Systems Engineering: Working proficiency in modern C++ for high-performance simulation integration.
The expected base salary range for this full-time position across US locations is listed below. Actual starting pay will be based on job-related factors, including exact work location, experience, relevant training and education, and skill level. Your recruiter can share more about the specific salary range for the role location or, if the role can be performed remote, the specific salary range for your preferred location, during the hiring process.
