Bagel Labs is a physical AI research lab building the model layer for autonomous robots. Our models learn how the world changes and how a robot should act within it. WorldDiT brings world modeling and action generation into one architecture. Paris shows how independently trained diffusion experts can work together as one model. We are bringing these lines of research together for physical AI.
We care more about research judgment and the quality of your work than conventional credentials. If you have a clear point of view on how models should connect prediction and action, we want to hear from you.
Role Overview
You will work on the core research problems behind models that predict what happens next and generate useful actions. The work spans action-conditioned world modeling, diffusion and flow models, latent dynamics, long-horizon rollouts, and embodied generalization. You will help extend WorldDiT and explore how Paris-style specialization and composition can improve physical AI models.
What You'll Do
- Build and study models that jointly represent future world states and actions.
- Improve action conditioning, temporal consistency, and long-horizon rollout quality.
- Test how models behave across tasks, objects, scenes, and embodiments.
- Explore specialist models, routing, and composition for physical AI.
- Design evaluations with clear baselines, useful ablations, and splits that prevent leakage across episodes.
- Connect prediction quality to policy behavior instead of treating an isolated metric as the result.
- Make experiments reproducible and explain what worked, what failed, and what should happen next.
Who You Might Be
You are a researcher who likes problems where prediction, action, and physical state meet. Your background may be in world models, robot learning, video generation, diffusion or flow models, simulation, representation learning, imitation learning, or reinforcement learning. You design experiments that settle questions. You care as much about failure cases as headline results.
Desired Skills
- Strong research judgment in world models, robot learning, diffusion or flow models, simulation, representation learning, imitation learning, or reinforcement learning.
- Experience building and evaluating modern machine learning models.
- The ability to design clean experiments with strong baselines, useful ablations, and clear failure analysis.
- Strong Python and experience with a modern machine learning framework.
- Clear technical writing and the ability to explain why a result matters.
What We Offer
- Competitive compensation and meaningful equity.
- Direct ownership of core physical AI research.
- A small technical team that moves quickly from an idea to a serious experiment.
- Paid travel to leading machine learning and robotics conferences.