(Yicai) March 11 -- AgiBot released its first general-purpose embodied artificial intelligence large language model yesterday, which allows robots to learn from videos and real-world scenarios and which the Chinese robotics startup claims is better than other such models on the market.

The Genie Operator-1 boasts Vision-Language-Latent-Action architecture, which is composed of a Multimodal Large Model and Mixture of Experts, and which enables robots to learn from human-generated videos, the Shanghai-based company said yesterday.

The chatbot can also keep evolving by learning from real-world issues it encounters, it said. And it can be easily transferred between different types of robots and can quickly adapt to new robot bodies.

When testing the GO-1 model on five tasks of varying complexity, it significantly outperformed existing top models, with the average success rate jumping 32 percentage points to 78 percent, AgiBot said.

Just a day earlier, Shanghai State-Owned Capital Investment linked arms with AgiBot to promote breakthroughs in and the large-scale application of embodied AI technology, with the aim of building Shanghai into the world’s leading smart robot hub.

AgiBot rolled out its 1,000th embodied AI robot in January, and it has recently started to mass produce its latest humanoid robot, the Expedition A2. The firm was founded in February 2023 by Peng Zhihui, a former member of tech giant Huawei Technologies’ "Huawei Talented Youth Program" and has a large number of social media followers.

Humanoid robots marked their first year of commercialization in 2024, with more than 2,000 units shipped globally, according to CITIC Securities. This year, the number is expected to increase significantly and the industry could see explosive growth.

Editor: Kim Taylor