NVIDIA researchers have developed an AI agent, called Eureka, that is capable of teaching robots complex skills such as training a robotic hand to perform rapid pen-spinning tricks. This agent autonomously writes reward algorithms to train bots, whether it be opening drawers / cabinets or even manipulating a pair of scissors.
Eureka leverages both GPT-4 LLM and generative AI to write software code that rewards robots for reinforcement learning. It uses a GPU-accelerated simulation in Isaac Gym to quickly evaluate the quality of large batches of reward candidates for more efficient training. The AI agent then constructs a summary of the key stats from the training results and instructs the LLM to improve its generation of reward functions. This means that the AI is always self-improving regardless of the robot its teaching — quadruped, bipedal, quadrotor, dexterous hands, cobot arms, etc.
- NVIDIA Ampere architecture, 2nd Gen Ray Tracing Cores, 3rd Gen Tensor Cores
- 12GB 192-bit GDDR6, 15 Gbps, PCIE 4.0; Boost Clock 1807 MHz
- IceStorm 2.0 Cooling, Active Fan Control, Freeze Fan Stop, Metal Backplate
Reinforcement learning has enabled impressive wins over the last decade, yet many challenges still exist, such as reward design, which remains a trial-and-error process. Eureka is a first step toward developing new algorithms that integrate generative and reinforcement learning methods to solve hard tasks,” said Anima Anandkumar, senior director of AI research at NVIDIA.
