Chinese company AgiBot, previously noted in the press for a Guinness World Record, has introduced a new embodied artificial intelligence called Go-3 Preview. The system is designed to help robots learn continuously, acquire new skills, and adapt to previously unseen tasks.
Architecture and benchmark performance
Go-3 Preview runs on the World Reasoning Action Model (WRAM) architecture, which integrates task reasoning, outcome prediction and action generation to support complex decision-making and action planning.
AgiBot reports that the model was evaluated on three benchmarks specifically created for embodied AI: RoboDojo, RoboTwin and RoboColiseum. According to the company, Go-3 Preview achieved the highest scores to date on all three tests.
Training pipeline and continuous learning
AgiBot says it intends to release the model as open source by the end of 2026 to enable broader use.
The Go-3 Preview is built to support continuous learning across pre-training, post-training and deployment phases. The training approach combines two data modalities: videos recorded from a human point of view that show how people interact with objects, and robot motion trajectories that link visual observations to physical actions and their outcomes. By increasing the model’s capacity to learn from these varied experience sources, AgiBot aims to convert them into skills that robots can apply to previously unseen tasks.
During pre-training, the model was conditioned to perform unfamiliar tasks across diverse environments, such as kitchens or multi-robot workspaces. The company states that Go-3 Preview can execute complex, dexterity-demanding tasks without requiring per-task fine-tuning for every new activity.
Genie Evolver: a distributed training platform
For the second training phase, AgiBot developed a separate system called Genie Evolver. Genie Evolver is a distributed training platform that aggregates data from roughly 100 robots along with physical simulations and world models that represent the physical environment. The platform enables robots to practice tasks, self-evaluate performance and improve through repeated attempts.
AgiBot reports that while the system initially showed 28.2% effectiveness, performance stabilized at 85.7% after five days of operation.
Why this matters
If AgiBot follows through on making Go-3 Preview open source by the end of 2026, the model could accelerate the emergence of robots capable of handling multiple workflows in manufacturing and other real-world settings. The combination of human-view video data, robot trajectory data and an aggregated distributed training platform offers a pathway for embodied AI systems to learn from real experiences and generalize to new tasks.
AgiBot’s claims — use of the WRAM architecture, top results on three embodied-AI benchmarks and the rapid improvement reported from Genie Evolver — indicate progress toward practical embodied AI. The broader impact will become clearer after an open-source release and wider industrial validation.
