Anyone weighing AGIBOT’s new GO-3 Preview robot model should know two things before trusting its headline claim: AGIBOT says it ranks first on three benchmarks, and the company supplied those rankings itself. AGIBOT unveiled the embodied AI model on Oct. 29 at its Innovation Day in Shanghai, and a PR Newswire release dated Nov. 3 repeated the claim of first place on RoboDojo, RoboTwin and RoboColiseum. GO-3 is designed for AGIBOT robots, so this article also covers the security and privacy exposure that comes with the model.

The release describes a system that learns across pre-training, post-training and deployment, and it promises open-source release by end of 2026. GO-3 Preview not yet open-sourced, so no weights or code are public.

WRAM Predicts What Happens Next, Then Chooses Matching Actions

GO-3 Preview is built on WRAM, short for World Reasoning Action Model, which chains three steps: understand the task, predict how the scene will change, and generate the action that produces that change.

A typical vision-language-action (VLA) model takes a camera image and text instruction and outputs motor commands directly. WRAM inserts an explicit forecast between seeing and acting. According to AGIBOT’s three core designs, CoAE is a control-oriented autoencoder, SVP is a single-step visual predictor and KASO is a selection step used in training.

CoAE extracts object position, shape and relationships from the camera image. SVP then predicts visual states for several future moments from the current view and the task in a single forward pass, and an inverse dynamics model combines that forecast with the robot’s own state to produce executable actions.

KASO Discards Predicted Futures That Conflict With Real Actions

Physical tasks have many valid solutions. A robot can approach a cup from the left or the right, so a model that predicts one future may disagree with the action a demonstrator actually took. AGIBOT’s release frames KASO as choosing predictions consistent with actions taken during training, which in practice means setting aside forecasts that conflict with what the demonstrator did.

The mechanism is described only in company materials. PingWest reported that AGIBOT has published no technical report for the model, so outsiders cannot yet check how the filtering works or how much it contributes.

Human First-Person Video Supplies About 90% Of Pre-Training Data

AGIBOT says GO-3 Preview learns from human first-person video, which the ai-bot.cn product page puts at about 90% of pre-training data, how physical states change and how objects are manipulated. Human body poses do not map directly onto robot joints, so the model pairs that video with robot trajectories, which supply the motor side. NVIDIA’s GR00T N1 also mixes human videos with robot data, so the data strategy is not unique to AGIBOT.

Physical RSI Means Recursive Self-Improvement, Not Human Feedback Training

After pre-training, AGIBOT’s Genie Evolver system combines a fleet of 100 physical robots with physics simulation and world models for post-training. A mechanism called Collective Experience Learning pools successes and failures across the fleet. AGIBOT reports that a phone-packaging task rose from 50.8% to 98.3% success as more robots took part at equal training time.

After deployment, an agent framework called Genie Master plans steps and checks results and resumes after interruptions, and what AGIBOT calls Physical Recursive Self-Improvement feeds real successes, failures and recoveries back into training. That usage is narrower than the software idea of an AI designing a more capable successor. Interesting Engineering reported, on AGIBOT’s account, that a G2 robot’s convenience-store workflow rose from 28.6% on day one to 85.7% after five days, and no outside party has confirmed it. Humanoids Daily noted that AGIBOT has not published evaluation details needed to judge how much of the improvement in its convenience-store demonstration was autonomous.

Do Three Leaderboard Wins Show GO-3 Works In Real Homes?

Not by themselves. Coverage of the launch treats the rankings as vendor-reported until others reproduce them. The three benchmarks also differ in what they test.

• RoboDojo, from the University of Hong Kong, combines 42 simulation tasks and 18 real-world tasks and has integrated 30 policies. When HKU announced it on Aug. 12, the top model scored only 8.80% in simulation and 12.8% in real-world testing, against 76.03% and 100% for human experts. Those are launch-era numbers for the best model then, not GO-3 scores. PingWest reported that GO-3 Nano Preview ranked first with an average score of 61.90 and an average success rate of 56.42%. Nano is a lighter variant, not the full Preview, and the two sets of numbers should not be compared directly.
• RoboTwin is a bimanual benchmark whose version 2.0 covers 50 bimanual tasks with domain randomization. Letsdatascience describes the GO-3 result as the highest overall score on RoboTwin, not an uncontested first place.
• RoboColiseum offers 78 simulation tasks in four sub-lists: instruction following, spatial understanding, disturbance adaptation and general operation. Project lead Wu Mo said simulation scores correlate 89.5% with real deployment and that the sim-to-real difference is under 10%. 36Kr did not independently verify those figures, and the 50 to 100 controlled experiments per task covered more than a dozen tasks, not all 78.

A check of the RoboDojo leaderboard page for this article returned no visible entries, so GO-3’s scores could not be confirmed there. The wider field shows why that matters. Stanford’s 2026 AI Index reported that the top method reached 89.4% on the simulated RLBench benchmark, while the best team on BEHAVIOR-1K, built around 1,000 household activities, reached 12.4% full-task success.

Open Weights, Real-Robot Data And Independent Audits Remain Missing

AGIBOT has promised to open-source GO-3 Preview by the end of 2026, but it has stated no exact date or license and has not said which training resources will accompany the release. Until weights and an evaluation protocol are public, outside labs cannot reproduce any ranking. International developers should expect friction that a leaderboard does not capture: a preview label, a Nano variant and no published technical report.

China’s National Intelligence Law Applies To Data From Robots Running GO-3

GO-3 is software, but it is designed for AGIBOT robots, which are made by a Shanghai-based company. Under Article 7 of China’s National Intelligence Law (2017), organizations and citizens must cooperate with national intelligence work in accordance with law, and Article 14 lets intelligence bodies request that cooperation. TechTimes examined that obligation in July. How far it compels a company to hand over robot sensor data is contested: legal analyst Jeremy Daum argues it is far from clear Article 7 was meant to require active information gathering, and that the law attaches no penalty for simply failing to cooperate.

China’s Data Security Law (2021) extends liability to overseas data handling when it harms national security, the public interest or the lawful rights of Chinese citizens or organizations. The Cybersecurity Law (2017) took effect Jan. 1, 2026 in amended form, with higher fines and wider reach over overseas conduct that harms China’s cybersecurity.

Data at stake. A deployed robot that sees, hears and navigates can generate camera video, audio, location traces and operational logs.

Audit status. No independent security audit of AGIBOT hardware or GO-3 turned up in the research for this article. That is a gap, not a clean bill of health. A separate security study of Unitree’s G1 described the humanoid sending telemetry to external servers every 300 seconds, which shows why audits matter, but it says nothing about AGIBOT.

US rules. On July 28 the FCC added foreign-produced advanced robotic devices to its Covered List. Devices that already hold FCC equipment authorization are not affected, and foreign producers can seek conditional approval from the Defense Department. The House passed its fiscal 2027 defense bill on July 22 with Section 163 barring Pentagon purchase of humanoid robots tied to foreign adversaries; it is a House-passed proposal, not law, and covers only Defense Department acquisition. A separate bill, H.R. 9129 was introduced June 3 and referred to the House Energy and Commerce Committee. Check current status before any purchase.

Practical steps. Segment robot networks, monitor outbound traffic, disable cloud sync of sensor data where possible, and review procurement rules. None of these removes the structural legal exposure.

Evaluate GO-3 On Independent Real-Robot Evidence, Legal Exposure And Release Terms

GO-3 Preview offers a concrete technical idea: forecast the next visual state, then pick the action that produces it. Its three first-place claims are AGIBOT’s, at least one rests on a lighter model, and the sim-to-real gap makes any leaderboard a weak guide to household reliability. Buyers and developers should wait for released weights, independent real-robot results and a security audit, and weigh Chinese data-access law as a condition of ownership whose reach is still debated.

Frequently Asked Questions

What is AGIBOT GO-3 Preview?

It is an embodied foundation model unveiled Oct. 29 that predicts future visual states and generates matching robot actions, using the WRAM architecture described in AGIBOT’s PR Newswire announcement.

Is AGIBOT GO-3 open source?

Not yet. AGIBOT says it intends to fully open-source GO-3 Preview by the end of 2026.

Are robot benchmark rankings reliable evidence of real-world performance?

Not alone. Stanford’s 2026 AI Index showed 89.4% simulated success on one benchmark against 12.4% full-task success on a household-activity benchmark.

Does China’s data law reach robots that run GO-3?

AGIBOT is headquartered in Shanghai, and the National Intelligence Law obliges Chinese organizations to support state intelligence work, though scholars debate how far that duty compels data handover.