Large language models (LLMs) can accelerate the training of robotics systems in super-human ways according to researchers at Nvidia, University of Pennsylvania, and the University of Texas.
The study introduces DrEureka, a technique that creates reward functions and randomisation distributions for robotics systems with a high-level task description.
DrEureka is a faster and more efficient solution for transferring learned policies from simulated environments to the real world.
LLMs can combine their vast world knowledge and reasoning capabilities with the physics engines of virtual simulators to learn complex low-level skills.
DrEureka uses a language-model driven pipeline for sim-to-real transfer with minimal human intervention.
Policies trained using DrEureka outperform the classic human-designed systems for quadruped locomotion by 34% and 20% in forward velocity and distance travelled, across various real-world evaluation terrains.
Best policy trained by DrEureka for robotic hands performed 300% more cube rotations than human-developed policies.
The safety instruction included in the task description plays an important role in ensuring that the LLM generates logical instructions that transfer to the real world.
DrEureka demonstrates the potential of accelerating robot learning research by using foundation models to automate the difficult design aspects of low-level skill learning.