What the role involves
One Robot builds task-specific world models and an evaluation platform for robot manipulation policies.
Training end-to-end policies for robots is vibes-based today. Teams collect data, train, deploy on a real robot, find out what fails, collect more, retry. We replace the trial-and-error with rigorous validation that tells you where your policy will fail and what data to collect to fix it.
Robotics can't industrialize without an evaluation layer. We're building it.
We're based in San Francisco, backed by Accel, YC, several exited founders, and engineering leaders at leading AI companies. We're small and deliberately so — everyone owns a wide surface area and moves fast.
This internship is for people who want to work on real problems with real robots, not toy datasets. You'll embed directly with the technical team and contribute to active research across world models, policy training, and evaluation.
What you'll do
- Contribute to training runs: Work alongside founding engineers on world model or eval model experiments end-to-end
- Run real-robot evaluations: Collect demonstration data, run policies on physical hardware, and document failure modes
- Build tooling: Write Python and PyTorch code that improves our data engine, training pipelines, or eval infrastructure
- Close the sim-to-real gap: Run experiments that test how well simulation predicts real-robot behavior
Requirements
- Strong coding in Python and PyTorch
- Currently enrolled in a BS, MS, or PhD program in ML, robotics, computer vision, or a related field
- Hands-on experience training or fine-tuning a generative model, VLM, or policy (coursework or research counts)
- Ability to work in-person in San Francisco
Nice to have
- Experience with real robot hardware or simulation environments (Isaac, MuJoCo, etc.)
- Prior research in manipulation, 3D vision, or model evaluation
What they ask for
About One Robot
One Robot builds simulation environments that are realistic to see and realistic to interact with, so robotics teams can train and evaluate robot policies without being bottlenecked by robot time. Today, improving a VLA often means more real-world hours: setting up the scene, running trials, resetting, and repeating. This loop is slow, expensive, and hard to scale. For example, material handling and manufacturing assembly tasks, models need far more training and evaluation data than teams can collect in the real world. We use task-specific data to build world model-based simulation environments for hard manipulation tasks (for example, textiles and box folding). These environments help teams run more training and evals, find failure modes faster, and accelerate iteration on action policies with less dependence on real-world data collection and robot availability.
Full One Robot profile