SHANGHAI, China, October 29, 2026 — AGIBOT has launched its latest innovation, the GO-3 Preview, an advanced embodied foundation model designed to enhance learning and adaptability in robotics. This model integrates pre-training, post-training, and post-deployment learning, aiming to develop robots with general-purpose capabilities that can evolve through real-world experience.
The GO-3 Preview has achieved first-place rankings in three leading embodied AI benchmarks: RoboDojo, RoboTwin, and RoboColiseum, with plans to be fully open-sourced by the end of 2026. Built on AGIBOT’s proprietary World Reasoning Action Model (WRAM), it combines task understanding, future-state prediction, and action generation, enabling robots to generalize across tasks and improve their manipulation skills.
AGIBOT’s innovative approach, termed “full-lifecycle scaling,” allows robots to refine their abilities not just during the initial training phases but throughout their operational life, leveraging real-world interactions to enhance performance. The company aims to enable robots to learn not only how to perform specific tasks but also to grow more competent through ongoing task execution.
AGIBOT highlighted the importance of transferring capabilities across contexts. For instance, in a simple task like picking up a cup, robots must adapt to varied object types and situational contexts rather than memorizing fixed movements. GO-3 Preview addresses this with WRAM’s ability to scale both data and model capabilities, facilitating quicker adaptation to new tasks.
In addition to pre-training advancements, GO-3 Preview integrates Genie Evolver, a hybrid post-training system, which combines physical robots and virtual environments to enhance learning. This system employs Collective Experience Learning (CEL) to consolidate the knowledge acquired by a fleet of robots, allowing successful strategies to be shared and optimized across the group.
Post-deployment, AGIBOT has developed Genie Master to enable robots to continue learning through their operational experiences. This framework facilitates planning, execution, and feedback, allowing robots to self-improve as they encounter new challenges without requiring human intervention.
AGIBOT’s recent demonstrations include using the Genie G2 robot in a real convenience store, showcasing its capacity for autonomous improvement by learning from actual operational data rather than relying on pre-programmed behaviors.
The company’s past models, including GO-1 and GO-2, laid the groundwork for the GO-3 Preview’s full lifecycle learning strategy, reinforcing AGIBOT’s commitment to advancing robotic capabilities through continuous learning and adaptation.
For more details about AGIBOT and its innovations, visit AGIBOT.com.
