To that extent, the RAI Institute collaboration focuses on reinforcement learning, a trial-and-error-based method. Simulations now accelerate this traditionally time-intensive process by enabling multiple learning tasks to run concurrently.
Earlier, Chinese robotics company AgiBot introduced the largest humanoid manipulation dataset to date, designed to enhance AI robot training. Called AgiBot World Alpha, the dataset includes over one million trajectories from 100 robots across 100+ scenarios in five domains, focusing on manipulation, tool use, and collaboration.
Unitree also recently released an open-source full-body dataset, enabling its H1, H1-2, and G-1 humanoid robots to move more naturally, even dance. Captured using LAFAN1 motion capture, the data is compatible with all of Unitree’s main robot units.