AGIBOT Unveils GO-3 Preview, an Embodied Foundation Model That Keeps Learning After Deployment
The WRAM-based model tops three embodied AI benchmarks and will be open-sourced by the end of 2026.. AGIBOT unveiled GO-3 Preview, an embodied foundation model built on its WRAM architecture that spans pre-training, post-training and post-deployment learning.

AGIBOT has unveiled GO-3 Preview, an embodied foundation model it says learns across the entire life of a robot rather than only before it leaves the factory. Announced, the model is built on the company's WRAM architecture, which AGIBOT describes as a world reasoning action model that joins task understanding, prediction of future states and action generation in one system. The company says GO-3 Preview placed first on three embodied AI benchmarks, RoboDojo, RoboTwin and RoboColiseum, and will be fully open-sourced by the end of 2026.
The architecture rests on three pieces AGIBOT names CoAE, SVP and KASO. It says CoAE extracts visual features geared toward control, keeping track of where objects are and how they relate in space. SVP predicts a full future state in one step from what the robot sees and the task it was given, and the action module then works out how to move from the present toward that predicted outcome. KASO picks the predicted futures that best match the actions the robot actually took during training, which AGIBOT says keeps prediction and action generation aligned.
On the data side, AGIBOT says the model learns from human first-person video and from robot trajectories, pairing examples of people handling objects with records of what a robot observed, did and what happened next. The company argues the point is not to memorize more examples but to get more out of each one when a robot meets something unfamiliar. It says GO-3 Preview can take on new tasks with no task-specific fine-tuning on a physical robot, and can adapt further from a small number of demonstrations.
For the post-training stage, AGIBOT describes Genie Evolver, a distributed system that combines a fleet of 100 physical robots with physics simulation and world models. It says a method called Collective Experience Learning pools what the fleet does, so successful trajectories become shared examples and the context around failures shows where to improve. Simulation lets the robots repeat scenarios under controlled conditions, while world models let the system revisit decision points and test alternative actions before trying them on real hardware.
AGIBOT says the model then keeps learning after deployment through Genie Master, its embodied agent framework, which uses an Embodied Execution Protocol covering capabilities, guidance, feedback and continuity. The company says the system records successes and failures, consolidates reusable knowledge into long-term memory at the end of each day, and turns execution experience into training data for better action policies. It demonstrated the approach with a Genie G2 robot working in a convenience store, where it says the robot diagnosed its own task failures and improved without human demonstrations or manual tuning.
The company frames this as physical recursive self-improvement in a real business setting rather than a laboratory exercise, and as the next step from the ViLLA framework of GO-1 and the action chain-of-thought approach of GO-2. What remains unproven is how much of the convenience-store gain carries over to other sites and tasks, and how the open-source release at the end of 2026 will hold up when outside teams run it on their own robots.
Source: AGIBOT.
InteliToys is editorial. We do not sell AGIBOT GO-3 Preview embodied foundation model. Watch the source tape in the branded lightbox below.


