Finalist
Soren
Training robots for high-stakes tasks, such as surgery, is bottlenecked by data.
Core trick
Generates parametric surgical simulation scenes with task-specific reward functions and trains PPO policies on failure cases.
Description
Training robots for high-stakes tasks, such as surgery, is bottlenecked by data. Real surgical data is scarce, hard to label, and too risky to generate through trial and error. This gap is even worse for rare complications and edge cases, which are exactly the scenarios where robotic precision matters most. This slows the development of safe autonomous systems and limits their potential to improve precision and expand access to high-quality surgical care.
Our project addresses this gap by creating a scalable source of realistic surgical training data. We developed a reinforcement learning environment-generation harness that leverages Astra’s spatial intelligence to construct high-fidelity surgical scenes with variable anatomy, deformable tissue behavior, procedural conditions, and failure cases.
Previously, environments of this quality were difficult to generate at scale, particularly with reliable, task-specific reward signals. Astra’s strong spatial reasoning makes both possible, enabling robotic systems to train and validate on complex, high-risk scenarios safely in simulation. By bringing the capabilities of frontier AI models into surgical robotics, we hope to accelerate the path toward safer autonomous systems and, ultimately, make high-quality surgical care more precise and accessible to more people.
README excerpt
# Soren Soren is a simulation workspace for finding robot-policy failures, correcting them, and testing whether post-training improves the result. It combines native MuJoCo physics, an interactive Three.js frontend, scenario planning, operator demonstrations, and PPO training. The workflow is simple: **build an experiment -> evaluate the starting policy -> inspect failures -> adjust rewards or demonstrate a correction -> train -> compare matched runs**. ## The Problem Training robots for high-stakes tasks, such as surgery, is bottlenecked by data. Real surgical data is scarce, hard to label, and too risky to generate through trial and error: a robot cannot learn proper suturing technique by practicing on real patients. This gap is especially acute for rare complications and edge cases, where reliable performance matters most. Without diverse, representative training scenarios, it is difficult to design useful rewards and evaluate whether a policy can handle failure conditions. The bottleneck also includes tooling. For every new task, someone must build a simulation environment, model the objects involved, and design a reward function, often through painstaking trial and error. That work is expensive even for simple tasks. In surgery, the scenarios most important to cover can also be the hardest and riskiest to collect real data on. ## Our Solution: An Interface for Instant RL Environments Our product vision is a pipeline where you describe a robotics task in plain lang