128 Large Language Models and Generative AI Should Accelerate The Progress In Robotics Trained by GPT-4 to perform robotics tasks, a neural network performed better than human expert coders on 83% of tasks, with the margin of improvement averaging 52%. Large Language Models (LLMs) enable text-based training, validation, and self-explanations, which should facilitate regulatory approval. Multimodal models can train autonomous vehicles with images and text, which could result in better performance. Generative AI can train and validate autonomous vehicle safety through simulation. IS LLM-Driven Reinforcement Learning Outperforms Expert Human Coders Task Legend: X A T Across Various Robotics Tasks, Environments, And Morphologies Task 1: To open the cabinet door O Task 2: To make the hand spin the object B RO Eureka (LLM-based reward design with little manual input, zero-shot rewards) toward a target. L2R (LLM-based reward design with manual reward templates, few-shot examples) Task 3: To make the humanoid run as fast as Human possible. 12.65 Task 4: To make the ant run forward as fast as 12.64 2.07 2.06 possible. 1.66 Task 5: To make the shadow hand spin the 1.42 object toward a target. 1.24 1.09 Task 6: To make the quadruped follow 0.88 0.99 1.06 1.00 1.00 1.00 1.00 randomly chosen x, y and yaw target velocities. 0.56 Task 7: To make the quadcopter reach and hover near a fixed position. -2.04 -1.04 Task 8: To balance a pole upright on a cart. Human Normalized Score Task 9: To stabilize a ball on the table-top. Task 1 Task 2 Task 3 Task 4 Task 5 Task 6 Task 7 Task 8 Task 9 Note: Yaw is rotation along the vertical axis of an aircraft. Sources: ARK Investment Management LLC, 2024. This ARK analysis is based on a range of underlying data from external sources, including Ma et al. 2023 and Wayve 2023, which may be provided upon request. Forecasts are inherently limited and cannot be relied upon. For informational purposes only and should not be considered investment advice or a recommendation to buy, sell, or hold any particular security. Past performance is not indicative of future results.