AGIBOT GO-3 Preview
First on three benchmarks
AGIBOT says its GO-3 robot software is first on three tests, and on the two whose tables could be read it leads one by its own count and is second on the other.
Robots · Checked 13 Nov 2026 · 14 sources · 10 min

A video of a robot lifting a bottle off a table, passing it to its other hand and standing it in a box shows that the robot did that once. It doesn't show how many takes came before, or what happens when the bottle starts a hand's width to the left. The field relies instead on a duller kind of evidence: give the robot a job, have it try many times, and count how often it finishes.
A count only allows a comparison when everyone counts the same thing. So research groups publish fixed sets of jobs with fixed scoring rules, and keep a public table of how each entrant did. The set of jobs is called a benchmark and the table a leaderboard, or board. On 29 October 2026 the robotics company AGIBOT announced GO-3 Preview and said it ranks first on three benchmarks, RoboDojo, RoboTwin and RoboColiseum. The release gives no scores. What "first" amounts to has to be read off the tables, and two of the three could be read.
The thing being ranked is software. A program of this kind sits between a robot's cameras and its motors: it takes in what the cameras see and what the robot has been told to do, and puts out the movements the robot should make. Nobody writes those movements out by hand. The program learns them from examples, which is called training, and a program made that way is called a model. AGIBOT calls GO-3 Preview an "embodied foundation model". Embodied means it works through a body. A foundation model is trained once on a great deal of material and then adapted to one job after another. By AGIBOT's account the material was first-person video of people and recorded movements of robots.
The tables are built from counts. Suppose a robot is given the bottle job 100 times and stands the bottle in the box on 80 of them. Its success rate is 80 in 100, or 80%. Each attempt is called a trial. Many trials are needed because a robot's result is not fixed the way a sum is. The bottle starts in a slightly different place each time, and a grip that holds on one trial can slip on the next.
A success rate has a weakness. The examples a model learns a job from are recordings of the job being done properly, called demonstrations. A model shown the bottle job on one table under one lamp may have learned the job. Or it may have learned that table and that lamp, and be replaying motions that happen to work there. A high score in the same scene can't tell the two apart.
RoboTwin 2.0 tests for the difference. It is a benchmark of 50 two-arm tasks that can alter the scene around each one: the clutter on the table, the lighting, the background, the table's height and the wording of the instruction. A scene with nothing altered is called clean, and an altered one randomized. Each model on the board is trained on 50 clean demonstrations per task, then tested in 100 trials per task in clean scenes and again in randomized ones. For the bottle, randomized means a higher table, a different lamp, a mug in the way and the request worded differently. A model that memorized its motions would reach for the spot where the bottle used to be.
One table, sorted three ways
AGIBOT's row is not under the announced name. The public data files of the two boards that could be read both list it as "GO-3 Nano Preview", under team AGIBOT. AGIBOT's pages never say "Nano", and no source says whether the model that was tested is the model that was announced or a smaller variant. On RoboTwin 2.0 the entry, listed on 25 October, scores 88.54% in clean scenes and 81.48% in randomized ones. Take 81.48 from 88.54 and 7.06 is left, so altering the scenes takes about seven points off this model's score.
The two scores give three ways to order the entries. Sorted by the clean score the entry is sixth, behind a starred entry named PrismWAM at 91.84%. Sorted by the randomized score it is first. The board's default ranking averages the two. Here 88.54 and 81.48 add up to 170.02, and half of that is 85.01, which is first of 33 entries, ahead of that PrismWAM entry at 83.67. That entry is ahead in clean scenes and behind on the average, so it must lose more when the scene changes, and the default ranking counts that against it. AGIBOT's event release says the model "achieved the highest overall score on RoboTwin", and the board's default ranking bears that out.
RoboTwin 2.0 leaderboard
Sixth or first, depending on the sort
| GO-3 Nano Preview | PrismWAM | |
|---|---|---|
| Clean scenes | 88.54%, sixth | 91.84% |
| Randomized scenes | 81.48%, first | Behind |
| The average, the default ranking | 85.01, first of 33 | 83.67 |
The table also records who did the counting. Each row has a field named contributor, which the board defines as "team that produced the numbers", and on this row it reads AGIBOT. So the 88.54 and the 81.48 are AGIBOT's count of AGIBOT's own model. A count like that can be exactly right and still has a limit. Whoever runs the trials is the only one who saw them, and the party that counted is the party that gains from a high number.
The remedy is for someone else to run the trials again, which takes the model itself. A trained model is a very large set of numbers, called its weights, and the code around them does nothing without them. RoboTwin 2.0's listing policy requires "publicly released code, publicly released weights, and a technical report".
For this entry, the places checked did not have them. What is public is the connecting code that lets a testing toolkit named XPolicyLab run the model through a benchmark's tasks. On 13 November 2026 its instructions stated "Installation does not download model weights" and linked to no weights file, and AGIBOT's account on Hugging Face, a site where such files are commonly shared, listed six models, none of them GO-3. Whether the weights were released somewhere else is unknown.
Overtaken on 30 October
RoboDojo, the second board, differs in what is counted and, by its own rules, in who counts. Its protocol says scores "are computed by the official evaluation system" and not reported by participants, and its news log records the entry as "contributed by AGIBOT". Go back to the bottle job and its 20 failed trials. In some the robot might never get a grip. In others it might lift the bottle, pass it across and drop it at the box. A success rate treats them all alike. RoboDojo reports a second number beside it, and its paper defines the pair: "success rate measures binary task completion, while average score captures partial task progress". Binary means yes or no.
Suppose all 20 failed trials got the bottle as far as the box, and that a finished job earns 100 points and a dropped bottle 50. The 80 finished trials earn 8,000 points and the dropped ones 1,000, and 9,000 points over the hundred trials is an average score of 90, while the success rate stays where it was. Those marks are invented to show the principle, and RoboDojo's own scoring is defined in its paper.
In RoboDojo's data dated 3 November, the entry has a score of 61.9 and a success rate of 56.42%. For scale, the paper reports that human experts under the same protocol reached 76.03% success and a score of 80.42. So this entry completes a little over half of its trials where human experts complete about three in four.
The entry's numbers come from simulation. RoboDojo has 42 simulation tasks, each run 50 times, and 18 real-robot tasks, and the entry is absent from the real-robot board. In simulation the robot, the table and the bottle exist only inside a program that works out what happens when an arm moves. That makes a trial cheap to repeat and the same for every entrant. The price is that a simulated world is simpler than a real one, and a model that does well in it can do worse on a real machine.
A leaderboard also carries a date. AGIBOT's event release says the model "ranked first in both task success rate and process score on RoboDojo", where process score most plausibly means the average score, though no source states this. The news log dates AGIBOT's entry to 27 October and an entry called Q0-Preview, from Qoo Tech, to 30 October, the day after the announcement. The data file gives Q0-Preview a score of 74.2 and a success rate of 68.87%. Every other simulation entry scores lower than GO-3. So the entry probably led the simulation board on the day AGIBOT said so, and in the data dated 3 November it is second by score. On one dimension, which the board calls precision, it is still first of 66 entries, at 78.65.
RoboDojo also sets a deadline for the weights. Its protocol requires the tested copy of them, which it calls a checkpoint, to be released through XPolicyLab within three days of listing, and its committee reserves the right to remove entries that miss this. No removal had been reported by 13 November 2026.
The third table could not be read, because RoboColiseum's site draws its content by script. That leaves AGIBOT's statement that the model "ranked first in overall score on RoboColiseum" unchecked. What can be read is the project's repository, the public home of its code. It says that running the benchmark on one's own machine needs "a separate Genie Sim checkout", a copy of a simulator. The simulator is AGIBOT's: its own repository states "Genie Sim is the simulation platform from AgiBot" and directs users to compete on RoboColiseum.
A promise dated the end of 2026
AGIBOT's release says the model "will be fully open-sourced by the end of 2026". If the weights are published as promised, outside groups could run the same RoboTwin and RoboDojo evaluations through XPolicyLab. Until then, only AGIBOT and the boards' own systems can reproduce the scores. No independent replication of any GO-3 result was found.
What anyone may do with the weights once they are out depends on the licence attached to them, and AGIBOT has not named one. For the earlier GO-1 model, the repository states that "the data and code within this repo are under CC BY-NC-SA 4.0", a non-commercial licence. If GO-3 follows that precedent, commercial use would be restricted, but no source says it will.
Check it yourself
- Open AGIBOT's release and read the first paragraph. It confirms the three first-place claims, the date and the end-2026 promise, and shows no scores or licence.
- Open the RoboTwin 2.0 leaderboard and read "Submission & listing policy". It confirms that listed models must release weights.
- Open the RoboTwin 2.0 data file and search for "GO_3_Nano_preview". It confirms 88.54 and 81.48, contributor AGIBOT and the listing date.
- Open the RoboDojo data file and search for "GO-3 Nano Preview" and "Q0-Preview". It confirms 61.9 against 74.2, and GO-3's absence from the "realworld" list.
- Open the RoboDojo protocol and read "Open-Source Artifact for Verified Publication". It confirms the checkpoint rule and the three-day deadline.
- Read "Model Assets" in the instructions for the GO-3 Nano Preview connecting code. It lists the checkpoint files needed and links none.
- Read the evaluation sections of the RoboDojo paper. They define both scores and give the 76.03% figure for human experts.
Sources
- AGIBOT Unveils GO-3 Preview, Takes Top Spot on Three Embodied AI Benchmarks. AGIBOT, 29 October 2026.
- RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation. Tianxing Chen and others, arXiv, 22 June 2025.
- RoboTwin 2.0 Leaderboard. RoboTwin Team.
- leaderboard.generated.json, RoboDojo leaderboard data. RoboDojo, updated 3 November 2026.
- robotwin_leaderboard.json, RoboTwin 2.0 Leaderboard data. RoboTwin Team, updated 30 October 2026.
- AGIBOT Introduces Worken to Measure Productive Robot Work at AGIBOT Innovation Day 2026. AGIBOT, 29 October 2026.
- GO-3 Nano Preview, XPolicyLab policy README. XPolicyLab on GitHub.
- agibot-world (AgiBot World). Hugging Face.
- Evaluation Integrity and Anti-Gaming Protocols. RoboDojo.
- RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies. Tianxing Chen and 43 other authors, arXiv, 5 July 2026.
- RoboColiseum/RoboColiseum: robotic evaluation guides and AI-assistant skills. RoboColiseum on GitHub.
- AgibotTech/genie_sim: Simulation Platform from AgiBot. AGIBOT on GitHub.
- AGIBOT GO-3 Preview Links Robot Pre-Training, Practice and Learning on the Job. Humanoids Daily, 30 October 2026.
- OpenDriveLab/AgiBot-World. OpenDriveLab on GitHub.