Want to know if this job is worth applying to?
Zürich, Zürich, Schweiz/Suisse/Svizzera/Svizra
Onsite
Odyssey is an AI lab pioneering general world models: causal, multimodal systems that learn to predict and interact with the world over long horizons. This foundational technology promises to revolutionize robotics, science, healthcare, education, gaming, defense, and beyond.
Odyssey’s founders previously pioneered the most complex application of physical AI: self-driving cars. They’ve now brought together a world-class research team from DeepMind, Tesla, Waymo, Meta, Apple, and Wayve, who have made significant contributions to language models (DeepMind Gemini), video models (DeepMind Veo), world models (Wayve GAIA), and autonomous systems (Tesla FSD).
Odyssey has raised significant venture capital from GV, Amazon, AMD, EQT, NVIDIA, Natural Capital, In-Q-Tel, Elad Gil, Jeff Dean, Guillermo Rauch, Garry Tan, Kyle Vogt, and researchers from OpenAI, DeepMind, MSL, Recursive, and Thinking Machines.
We are looking for people with a deep interest in improving machine learning algorithms at the intersection of diffusion world models, robotics simulation, and model-based agents. You should be drawn to the bet: that a model learned from the breadth of the real world covers the scenes, sensors, and situations no hand-built simulator enumerates, and generalizes to the ones nobody thought to model. You will build and use world models as simulators, feature extractors, and training grounds for real robot policies and autonomous driving stacks, latent imagination, planning and search, model-based RL. Interactive, action-conditioned world models are a cutting-edge research area that is not yet mature, so you will be working at the cusp of what’s possible. You will get the hardest version of it: dense multi-agent traffic, long horizons, surround sensing, and safety cases nobody can collect enough real miles for, the problem our founders spent a decade on. Most new experiments here will fail; your focus will be on maximally learning from failed experiments to increase the chances of eventual success on real hardware.
Learn what makes interactive world models tick. How data, diffusion backbones, action conditioning, and downstream robot and driving policies interact.
Push on what makes a world model usable as a robotics simulator: rollout horizon, temporal and physical consistency, controllability, multi-view and multi-sensor coherence (surround cameras, depth, lidar, proprioception), and drift under the policy's own actions.
Implement state-of-the-art diffusion, model-based RL, and robot-learning algorithms, and design losses, reward functions, and reinforcement / preference-based fine-tuning for interactivity and policy performance.
Train policies inside the model, latent imagination, planning, search, and close the loop by measuring them on real robots and in closed-loop driving evaluation.
Run a high cadence of ablations across world model and policy, from architecture and feature taps to data mixes and conditioning schemes.
Build the bridge to real robots and vehicles: data pipelines, world model fine-tuning, and policy deployment onto Robots-aaS and design partner systems (remote and face-to-face).
Exploit the latest features on modern GPUs to increase training and inference efficiency.
Take ownership of the full ML stack, including core frameworks that Odyssey researchers and product engineers rely on.
A PhD (or equivalent research experience) plus 2+ years of relevant research or engineering experience, or 4+ years of software engineering experience with 2+ years of relevant ML work.
Significant hands-on experience with one or more of: model-based RL (e.g. the Dreamer lineage), learned simulators for robotics or driving, diffusion and video generative models, RL and planning, or vision-language-action (VLA) policies.
Comfortable reasoning about a policy and the model it acts in as one system, rather than owning only one side of the interface.
Track record of owning projects end to end.
Not shy to touch any stage of an ML pipeline, from data to real-robot and in-vehicle deployment; sim-to-real experience is a strong plus.
Autonomous driving, or another safety-critical embodied domain, is a plus.
Proficiency with PyTorch (or TF/JAX).
Highly experiment driven.
Flexible to work in-person in Zurich, London, or the Bay Area. Most of the team works in-person in Zurich at the moment.
