PhD Research Intern Learned EndtoEnd Driving and Actuation

Most autonomy stacks are built around a single major sensor, usually the camera, with every other sensor secondary. Zendar took a different path: an early-fusion camera and radar perception system, in bird's-eye view and in 3D, in which each sensor plays a major role, trained data-efficiently on a multi-million-sample labelled dataset we collected ourselves. We own the whole pipeline, sensor setup and calibration, learned perception, lane estimation and map integration, and that vertical integration lets us move fast on hard problems.

We are now turning to what comes after perception: action and actuation, using the full pipeline to actuate autonomously in the real world, for autonomous vehicles and robotics, in a learned, end-to-end way. A learned motion planner drives on that stack today; this role is about the research that takes it further.

The Perception and Machine Learning Team is looking for a final-year PhD student to join us as a Research Intern on learned planning and actuation, on-site in Berkeley, CA. This internship offers an opportunity for both the intern and the team to assess mutual fit for a potential full-time research position.

You will own one research direction on the production planner, working with the researchers who ship it, on real fleet data, and present your results to the people who decide what ships. What a successful planner means to us, and what we will measure you against:

• Understands what is missing: finds the mechanism behind a driving planner's failures and shows which gap matters before changing anything.
• Handles the long tail, never gets stuck and knows how to get itself unstuck: agile, it commits, launches, re-plans and recovers on its own.
• Collision- and comfort-aware, and generalizes across structured environments (roads, lanes, intersections) and unstructured ones (lots, yards, sites, robotics settings).
• Knows how to test and evaluate actuation and planning pipelines: open-loop, closed-loop and on-vehicle, with metrics that predict road behavior.
• Production quality, not a proof of concept: tested code, reproducible results, behavior that holds across checkpoints and scenes. This is our standard and our measure of success.

• Generative planning with flow matching: a flow-matching or diffusion policy over ego trajectories with few-step inference and explicit constraint guidance in the sampler.
• The long tail: data balancing and hard-case mining, counterfactual and generative scenario synthesis, or reasoning-augmented planners.
• Closed-loop evaluation and training that matches our sensors: counterfactual observations faithful to radar and camera features, so closed-loop metrics predict road behaviour.
• Post-training with reinforcement learning: rule-reward or closed-loop fine-tuning, parallel imitation + RL—and whether the gain survives a real sensor stack and a safety layer.
• Your own idea: tokenized planning with learned scoring, a world model as the planner's critic, uncertainty-aware ranking, self-play traffic.

Every change is evaluated closed-loop and open-loop against the current planner, failures examined and not just counted, and results presented to the team's leads. We support publication of non-proprietary findings.

• Final-year PhD student (or within 12 months of defending) in computer science, electrical engineering, robotics or a related field, available for an on-site internship in Berkeley.
• First-author work in motion forecasting or planning, generative models for sequential decision making (diffusion, flow matching), reinforcement or imitation learning for control, or closed-loop simulation and evaluation.
• Python and PyTorch: you can read a large model file, change the loss, and prove it did what you…