Robot Agent eXploration (RAX): Scaling Online Exploration for Self-Improving Robots

Researchers from Harvard, NVIDIA, UC Berkeley, MIT, and Stanford introduce RAX, a robot learning framework that enables continuous self-improvement through online exploration. RAX allows robots to acquire and reuse skills without human demonstrations, achieving significant performance gains across multiple benchmarks and real-world tasks. The system uses code-as-policy and skill accumulation, enabling robots to adapt to new environments and tasks without retraining or fine-tuning. RAX demonstrates average improvements of 37.4 percentage points over existing state-of-the-art models, with success rates reaching 98.5% on some tasks.