Coffee-Serving Butler Bot Gets Souped-up AI Model – CoffeeTalk

Chinese robotics company AgiBot has introduced the Genie Operator-1 (GO-1) AI model to improve humanoid robots’ ability to perform real-world tasks. GO-1 uses vision-language models to process vast image and video data, helping robots better understand human actions. Its planning and action algorithms enable robots to map out steps and execute movements efficiently. AgiBot released AgiBot World in late 2024, a large-scale robotics dataset with over 1 million trajectories across 217 tasks in five domains.

The Vision-Language-Latent-Action (ViLLA) framework is introduced to enhance robot learning by combining vision, language, and action modeling. The framework comprises two key components: a Vision-Language Model (VLM) and a Mixture of Experts (MoE). The VLM processes vast amounts of multimodal data from the internet to build scene understanding and language comprehension. The MoE has two parts: the Latent Planner, which learns general action patterns from various sources, including human actions and different robot embodiments, and the Action Expert, which is trained on over a million real-world robot demonstrations to refine movement execution.

GO-1 presents latent actions, which use past and present frames to help robots comprehend movement. By transferring knowledge from many data sources, this method increases robots’ adaptability. The Latent Planner uses a specific transformer model to forecast action sequences, while the Action Expert improves movement execution using a denoising technique to guarantee fluid and accurate movements.

In tests across five tasks, GO-1 outperformed state-of-the-art models, improving success rates from 46% to 78%. It demonstrated significant progress in tasks like replenishing beverages and pouring water. AgiBot claims that GO-1’s capacity to learn from both human and robot data enables it to adjust to novel tasks, function with various robots, and constantly advance in real-world scenarios.

Read More @ MSN