AgiBot launches GO-3 Preview embodied model, plans to open-source it by end of 2026

AgiBot launched GO-3 Preview on October 29, an embodied AI model that the company says can keep learning at scale across pre-training, post-training and after deployment. AgiBot plans to open-source the full model by the end of 2026, according to Pandaily.

Architecture and rankings

GO-3 Preview is built on what AgiBot calls WRAM, or World Reasoning Action Model. The architecture links task reasoning, prediction of what happens next, and action generation. In AgiBot's description, a robot can understand an instruction, anticipate the outcome and produce motions it can carry out.

The model has recently taken first place on three embodied AI leaderboards: RoboDojo, RoboTwin and RoboColiseum. Specific scores were not given. Parameter count, context window and pricing are not yet disclosed.

Two scaling lines

AgiBot says the model scales along two lines:

• Data: first-person human videos supply examples of objects being handled and tasks unfolding, while robot trajectories tie observations and actions to physical results.
• Model: stronger learning capacity turns this scattered experience into skills that transfer to new tasks.

Pre-training results

After pre-training, AgiBot says robots can complete unfamiliar tasks in homes, kitchens, multi-robot setups and dexterous manipulation without extra fine-tuning. With a few demonstrations, the model can learn long, delicate jobs such as making a jianbing, a Chinese savory crepe. After fine-tuning, and with an agent coordinating the steps, a robot can also assemble Lego pieces.

Post-training: Genie Evolver

For post-training, AgiBot built Genie Evolver, a distributed training system that combines about a hundred real robots with physics simulation and world models. Reported results:

• In tabletop curling, a robot surpassed the average tested human after four hours of online practice.
• In a block-tower game, a robot pulled out five blocks in a row while keeping the tower standing.

Deployment: Genie Master

After deployment, the Genie Master agent framework connects judgment, execution and feedback so robots can keep improving on the job. An AgiBot G2 robot working in a real convenience store located the causes of its own failures, adjusted its approach and stored what it learned, without human demonstrations or tuning.

Its success rate on the full task flow rose from 28.6 percent on the first day to a steady 85.7 percent after five days. On a task such as hanging small hooks, AgiBot says this method improved single-task success by 17.7 percentage points over conventional supervised fine-tuning, and the robot continues to learn online.

What this means

GO-3 Preview is presented as one model spanning three stages that are usually treated separately: pre-training on human video and robot trajectories, post-training with real robots plus simulation, and learning in the field. The convenience-store result is the most concrete evidence of the third stage, with a full task flow improving over five days without human demonstrations or tuning. It covers a single robot and task flow, so it does not show how far the approach generalizes.

The leaderboard first places on RoboDojo, RoboTwin and RoboColiseum are reported without scores, so the margins are unknown. The planned open-sourcing of the full model by the end of 2026 would let outside teams test AgiBot's claims on their own hardware and tasks. Until then, the performance claims come from AgiBot.