MIT’s AI trains robot dogs in virtual worlds

The Brief: MIT researchers unveiled an AI system called LucidSim that trains four-legged robots using generated imagery — achieving unprecedented real-world performance without ever seeing actual environments during training.

The details:

LucidSim combines physics simulations with AI-generated scenes to create diverse training environments for robotic learning. Robots trained in LucidSim’s artificial environments completed complex tasks like obstacle navigation and ball chasing with up to 88% accuracy. The platform uses ChatGPT to auto-generate thousands of scene descriptions, creating varied training scenarios with different weather and lighting conditions. Traditional training methods relying solely on human demonstration achieved only 15% success rates on the same tasks.