Case Study: Decart builds Oasis 3 for scalable robotic learning with Weights & Biases

Decart builds Oasis 3 with Weights & Biases to power 1,000x faster world models

Decart, a frontier AI research lab, faced the challenge of developing complex world models for robotic learning, which required managing numerous interdependent variables during training. Tracking experiments and maintaining provenance for datasets and model checkpoints was difficult using scattered logs and tribal knowledge. To build their flagship product, Oasis 3, they turned to Weights & Biases to bring order and control to their development workflow.

Using Weights & Biases Models as the backbone of their development, Decart tracked all training experiments in one place. The platform's Sweeps accelerated hyperparameter optimization, while Artifacts versioned datasets and models, providing a complete audit trail by default. This solution gave Decart the necessary control and provenance tracking, allowing them to move fast on releases while meeting the compliance requirements of their serious robotics partners. The integration of Weights & Biases was essential for building the first API-accessible world model capable of real-time, interactive simulation.

Decart