2027 Summer Intern, PhD, Machine Learning - Simulation & Closed-Loop Training

Waymo•Mountain View, CA

8h•Hybrid

About The Position

Waymo is hiring a PhD Machine Learning intern to work on simulation and closed-loop training for autonomous driving. The intern will design, implement, and scale reinforcement learning post-training and distillation algorithms for foundation world models in simulation and autonomous driving. Work may involve flow or diffusion policy optimization, off-policy importance sampling, and constrained multi-objective reinforcement learning. The role also involves developing learned reward models, realism critics, and discriminators in a closed-loop multi-agent simulator to align simulated agent behavior with human driving realism and safety. The intern will conduct rigorous closed-loop ablation studies on large-scale TPU clusters, using agentic auto-research workflows and fleet-scale driving behavior metrics to mitigate simulator exploitation and distribution shift. They will write clean, modular, well-tested JAX code and collaborate with researchers and engineers across simulation, world modeling, and autonomous driving teams. This is a hybrid onsite internship. The position is full-time, with an expected hourly rate of $85. Interns may participate in the company’s benefits programs, subject to eligibility requirements. Applications are accepted on a rolling basis until the role is filled.

Requirements

Currently pursuing a PhD in Computer Science, Machine Learning, Robotics, or a related technical field.
Anticipated graduation date between December 2027 and August 2028, with a return to a degree program for at least one academic term after the internship.
Strong tensor implementation and debugging skills in Python and JAX/Flax or PyTorch.
Hands-on experience training and modifying modern neural architectures such as Transformers, Diffusion/Flow models, or deep reinforcement learning networks.
Hands-on research experience in either reinforcement learning and post-training or generative world, sequence, or trajectory modeling, with working familiarity in the other area.
For reinforcement learning and post-training: experience with policy gradient methods, off-policy or offline RL, constrained MDPs, inverse RL, or learned reward and critic modeling.
For generative world, sequence, or trajectory modeling: experience with continuous flow matching/Rectified Flows, score-based diffusion policies, or spatiotemporal autoregressive world and trajectory models.

Nice To Haves

First-author publications at top-tier machine learning or robotics venues such as NeurIPS, ICML, ICLR, CVPR, CoRL, RSS, or ICRA, or equivalent, in RL post-training, flow/diffusion policy optimization, generative world models, or autonomous driving.
Experience with multi-agent closed-loop simulation, Sim2Real transfer, or reward-hacking mitigation, including tools or settings such as Waymax, WOMD, NuPlan, IsaacGym/MuJoCo, adversarial traffic generation, or discriminator stabilization.
Experience profiling and scaling distributed training workloads on TPUs or multi-node GPUs, including XLA compilation, vectorized batched rollouts, and memory optimization.
Familiarity with C++/Python interfaces.

Responsibilities

Design, implement, and scale reinforcement learning post-training and distillation algorithms for foundation world models in simulation and autonomous driving.
Develop learned reward models, realism critics, and discriminators in a closed-loop multi-agent simulator.
Run closed-loop ablation studies on large-scale TPU clusters using agentic auto-research workflows and fleet-scale driving behavior metrics.
Mitigate simulator exploitation and distribution shift.
Write clean, modular, and well-tested JAX code in core simulation and autonomous driving codebases.
Collaborate with researchers and engineers across simulation, world modeling, and autonomous driving teams.