← All model news

AgiBot's GO-3 Preview Claims First on Three Robot Leaderboards, Focuses on Post-Deployment Learning, Year-End Open Source Planned

According to an October 29, 2026 Xinzhiyuan report, AgiBot released the embodied model GO-3 Preview, claiming first place on three robot leaderboards—RoboDojo, RoboTwin and RoboColiseum—and linking pretraining, post-training and post-deployment learning into one pipeline; the company expects a full open-source release by year-end, with weights, license and access methods not yet announced.

Published Updated

What changed

According to the October 29, 2026 Xinzhiyuan report, AgiBot released the embodied model GO-3 Preview that day, claiming it ranks first on three robot leaderboards—RoboDojo, RoboTwin and RoboColiseum—and in RoboDojo ranks ahead of the general-purpose model GPT-6 Astra mentioned in the report. As described, the model's focus is not a single metric but connecting three training stages into one line: pretraining uses a world reasoning action architecture called WRAM, comprising a control-oriented vision encoder CoAE, a single-step visual planner SVP and knowledge-aligned selective optimization KASO, alongside the release of a million-scale real-robot dataset AGIBOT WORLD 2026; post-training is handled by Genie Evolver, which alternates trial and error across three engines—real robots, physics simulation and a world model—and introduces failure-guided experience retrieval fine-tuning FERT, using the pre-failure context to retrieve segments from a fleet-shared success experience pool for weighted training; the post-deployment stage is taken over by the Agent framework Genie Master, released on October 27, connecting upper-layer reasoning and lower-layer execution through an embodied execution protocol called EEP, with failure lessons either consolidated into long-term memory rules (Harness RSI) or fed back into policy weights (Policy RSI). The reported figures include: average success rate on five real tasks—flower arrangement, clothes folding, phone charging, stool assembly and phone packaging—rising from 81.0% to 97.3%, the number of failures in the multi-stage long-flow phone packaging task falling by 92%, clothes-folding shape-control success rising from 85.8% to 98.3%, and full-process success rising from 28.6% to 85.7%. AgiBot says it expects GO-3 Preview and others to be fully open-sourced by year-end, and plans to run a 24-hour convenience-store livestream test with six G2 robots on November 2.

What it means for developers

For developers, this kind of 'continued learning after deployment' claim changes the system shape, not the parameters of a single inference call. If you treat the model as a static checkpoint, you will still have to fill in task orchestration, failure recovery and human takeover yourself when deploying; under the reported approach, part of the model's capability comes from experience accumulated during operation, so data return and policy updates are part of the system, not something to address after launch. The evaluation frame also deserves a second look: the rankings on RoboDojo, RoboTwin and RoboColiseum are provided unilaterally by the vendor, and the report does not disclose task distribution, evaluation scripts or reproduction methods; treating them as a signal of whom to try first is safer than treating them as a selection conclusion. The open-source timeline also needs room for uncertainty—the report says only 'expected full open source by year-end,' does not say whether weights, training code or data will be released, and does not mention a license; open weights are not the same as an open-source license, and rights for commercial use, redistribution and derivative models can only be judged once the license text is available. In its current state, there are no downloadable weights and no public API or access method; in the short term it cannot be plugged into existing services, so only architecture-level preparation and metric-level checking are possible.

Integration and evaluation

Below are conditional judgments based on the report, not facts announced by AgiBot. Decision 1: whether to invest in a data-return and retraining pipeline for 'post-deployment learning.' The relevant change in the report is that Genie Master claims it can raise full-process success from 28.6% to 85.7% without a human in the loop, and improve a single task by a further 17.7%. There are two options: treat the policy as a frozen model and remedy only at the memory and orchestration layer; or make failure-trajectory collection, policy updates and canary replacement part of routine operations. The precondition that changes the choice is whether you can reliably record sensor data and execution results for a period before and after failures in your own scenarios, and have a rollback-capable update process—if you cannot, you can only capture the half of the benefit that comes from memory and orchestration. Evaluation criteria could be set this way: on a fixed held-out set of your own tasks, repeat the same batch enough times to first establish the natural fluctuation range of the success rate, then see whether one round of update clearly exceeds that range, rather than judging from a demo video. Decision 2: start zero-shot, or do few-shot fine-tuning first. The report says GO-3 emphasizes zero-shot generalization, requiring no fine-tuning for unseen scenes and objects, and that after fine-tuning it can grasp, align and insert Lego parts. The choice depends on two things: whether you can collect demonstration data on the target embodiment, and how much error the task tolerates. If the task is a long process where one wrong step fails the whole thing, launching directly zero-shot carries higher risk than doing small-scale fine-tuning first. Evaluation criteria: on the same held-out tasks, compare zero-shot and few-shot on success rate, number of human interventions per completion and completion time; expand the scope only if at least two of the three improve consistently. Decision 3: wait for these weights, or use existing open-weight options first. The report offers an open-source promise with uncertain timing, and no license or repository information. If your project milestone comes before the weight release, building the pipeline on existing downloadable open-weight models is lower risk. The test can be simplified into three preconditions: the weights are downloadable, the license permits your specific use, and the bundled inference scripts run successfully and produce reproducible evaluation results; if any one is missing, treat it as 'not ready.'

Limits and open questions

The leaderboard rankings, success rates and efficiency gains in the report all come from one-sided accounts by AgiBot and Xinzhiyuan; the article does not provide evaluation scripts, task-set composition or third-party reproduction, and some figures come from real tasks selected by the vendor. The benefits of post-deployment learning depend on Genie Master, EEP and the G2 robot body; whether they hold on other hardware is not addressed in the report. Open source is described only as 'expected full open source by year-end,' with no license, repository address or weight files, so it can neither be downloaded nor called now; the report also has no public API or pricing information, and any availability claim should be judged against AgiBot's official position. GO-3 Preview is not in KimiSeek's current model catalog; KimiSeek has not launched this version, and the release and availability discussed here refer to the vendor side. In addition, comparison objects such as GPT-6 Astra mentioned in the text are provided by the source, and their evaluation setups are also undisclosed.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

刚刚,国产模型拿下三大榜单第一!年底全面开源

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.