AgiBot presented GO-3 Preview on 29 October 2026 at its Innovation Day in Shanghai — an embodied AI system that learns in three phases: from first-person human video, from practice on a fleet of real robots, and after deployment, on the job. Results shown at the event put it first on three robotic manipulation benchmarks — RoboDojo, RoboTwin and RoboColiseum.
Key takeaways
- GO-3 Preview debuted on 29 October 2026 at AGIBOT Innovation Day in Shanghai
- Roughly 90% of the pre-training data comes from first-person video footage, not from robot-recorded demonstrations
- At equal training time, growing the fleet of learning robots lifted phone-packaging success from 50.8% to 98.3%
- In a multi-stage packaging test, failures across 120 trials dropped from 25 to 2
- First place on RoboDojo, the top overall score on RoboTwin and first place on RoboColiseum, per figures presented at the conference
Picture the following situation
A humanoid robot at a convenience store shelf has to hang a product on a hook. A price tag blocks the head camera's view, the attempt fails, and a conventional system would log an error and wait for a human.
But a question arises: what if the robot analyses its own failure and changes how it acts?
In AgiBot's example the planning layer rewinds to the failed attempt and rearranges the motion — the left hand moves closer, the wrist camera supplies a close-up, the right hand finishes the job. That method goes into memory, ready for similar tasks. AgiBot calls this post-deployment scaling.
What does "full chain" mean in GO-3 Preview?
Robot control models have so far split a model's life into training and work. Once training ended, the action policy?action policy: a learned rule that turns the robot's view and state into a concrete motion froze, and whatever the robot met later never returned to it.
GO-3 Preview stretches learning across three stages AgiBot calls full-chain scaling: pre-training on video, post-training in a mixed environment of real robots, physics simulation and a world model, then gathering experience at work. It succeeds GO-1 with its ViLLA architecture and GO-2 with its action chain-of-thought.
This is a simplified diagram of the process AgiBot described, not a full technical specification.
How does a robot learn from footage that contains no robot?
Roughly 90% of GO-3 Preview's pre-training data is first-person video recorded by people. It covers far more scenarios than robot demonstrations, but has one flaw — a human hand and a robot gripper differ, and body poses do not map directly onto joint angles.
The answer is the WRAM architecture, chaining three steps: understanding the task, predicting the future scene and generating the motion. The SVP module handles prediction, while the KASO mechanism keeps predictions consistent with the motion actually performed. Human footage thus teaches how the world changes, robot trajectories?trajectories: the record of a robot's joint positions over time, the exact path a motion followed which motion causes it.
AgiBot splits the outcome into two modes: zero-shot measures what the model does without task-specific fine-tuning, few-shot adds a few demonstrations for precision. The company claims several hundred tasks work out of the box. In a Lego demo the model grasped, aligned and pressed in parts, and on the X2 bipedal humanoid it opened doors holding an object in both hands.
What does practising on a hundred robots at once buy?
Every attempt on a real machine costs something: the scene must be reset, hardware wears out, collision risk rises. So Genie Evolver, GO-3 Preview's post-training stage, mixes three environments.
Real robots supply trustworthy physics. AgiBot runs this stage on a cluster of over a hundred machines, where the CEL mechanism pools experience from every unit — successful motion fragments feed a shared policy, and failure contexts are analysed.
At equal training time, expanding that fleet lifted phone-packaging success from 50.8% to 98.3%. In a longer, multi-stage version of the task, failures across 120 trials fell from 25 to 2.
Simulation adds cheap repetitions with randomised friction and object positions. The world model rewinds to the key decision point to check what a different force or angle would have produced. AgiBot notes that the world model can misjudge friction and collisions, so real-machine validation stays necessary.
What do these names mean?
AgiBot's materials carry several acronyms.
- WRAM (World Reasoning Action Model) — architecture combining task understanding, future-scene prediction and motion generation in one pass
- SVP (Single-step Visual Planner) — predicts several future images of the scene in one forward pass. A separate inverse dynamics model?inverse dynamics model: a model that works out which motion moves the scene from one state to the next turns them into motion
- KASO — a selection mechanism. It generates candidate futures and keeps the one best matching the robot's real motion
- CEL (Collective Experience Learning) — collective learning. The equivalent of a shared repository instead of a copy per machine
- Physical RSI — recursive self-improvement in the physical world: successes and failures at work become material for the next learning round
They exist together so the same run — successful or not — feeds task planning and motion-policy training at once.
Why isn't it enough to train a model once and ship it to every robot?
That is standard practice today and works well in closed scenarios. The problem starts where the environment changes faster than the model's release cycle.
A convenience store is a good example. Shelf layouts, packaging and the product range all shift. Fine-tuning on new data means collecting demonstrations, labelling them and training — weeks of work.
There is a second problem. Most robot failures are not a lack of skill but edge cases — an item hidden behind a tag, a bag wrapped around the gripper. Those cannot be enumerated in advance.
Hence Genie Master, an agent framework that plans the task, checks every step and, after an interruption, resumes from where it stopped rather than starting over. Decision and execution layers talk over the EEP protocol.
Is it working yet?
As a technology preview, not a generally available product. The name "Preview" is itself a vendor signal. AgiBot showed the model on 29 October 2026 with no release date or access terms.
The deployment piece is furthest along. Per AgiBot, the Genie G2 robot running Genie Master is in validation in a working store, improving at deli packing and restocking without manual demonstrations. The company calls this the industry's first test of Physical RSI outside the lab — a vendor claim, not independently verified.
Benchmark results also come from a conference presentation, not a public leaderboard. The three test suites measure different things: an evaluation spanning simulation and real hardware, bimanual manipulation, and a general embodied AI ranking.
| Benchmark | What it covers | GO-3 Preview result |
|---|---|---|
| RoboDojo | Evaluation in simulation and on real hardware | First place |
| RoboTwin 2.0 | 50 bimanual manipulation tasks | Top overall score |
| RoboColiseum | Embodied AI ranking | First place |
Table: Results per figures presented at the conference, not from a public leaderboard.
Why does this matter?
Shifting the centre of gravity from training-set size to the post-deployment feedback loop changes the economics of service robotics. If a fleet improves itself from its own work, entering a new site drops from an integration project to a simple rollout. The measure of competitive advantage shifts too — what counts is robots actually working, not robots sold.
What next?
- AgiBot gave no release date for the full GO-3 — "Preview" status holds until further notice
- Genie G2 validation in the store is ongoing and should feed the next training round
- The same Innovation Day also brought the Worken framework, the WITA-Omni model, the AHA3 platform and the G2 Max robot
Sources
- 机器之心 (Jiqizhixin) — 刚刚，机器人学习进入「全链路Scaling」时代
- PR Newswire — AGIBOT Introduces Worken to Measure Productive Robot Work at AGIBOT Innovation Day 2026
- RobotHOT — AGIBOT launches GO-3 Preview embodied model for full-chain scalable learning
- RoboDojo — A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies
- RoboTwin — RoboTwin 2.0 Official Document
