DrEureka: Language Model Guided SIM-to-Real Transfer

The website discusses DrEureka, a novel approach for sim-to-real transfer utilizing Language Models (LLMs) to automate the design of reward functions and physics parameter randomization for robust robot skill acquisition. It achieves notable success without human intervention in tasks like quadruped Globe walking and dexterous cube rotation, as well as handling unexpected disturbances like kicking or deflating a yoga ball.

• DrEureka uses LLMs for sim-to-real tasks.
• The method requires only the physics simulation.
• DrEureka improves the design-tuning process.
• LLMs help build reward-aware physics priors.
• It successfully tackled novel robot tasks.