According to AgiBot, after only a few hours of training that combine real-world trials and learning through a world modelA learned model that represents an environment and predicts how it evolves in response to an agent’s actions.Read the glossary entry, a robot plays table curling with a higher win rate than ordinary human players.
For the refinement phase, AgiBot developed Genie Evolver. The system combines a fleet of 100 real robots, physics simulation and “world models”, software that predicts what may happen next.
AgiBot says it ran a demonstration with its Genie G2 robot in a real convenience store. Controlled by Genie Master, the software that organises its tasks, the robot reportedly found the causes of its failures and adjusted its approach while working.
Source
- Robotics & Automation News · roboticsandautomationnews.com