AGIBOT2026-10-30 · Ian Lapham
AGIBOT Leads Humanoid Shipments. On October 29 It Asked the Industry to Count Worken, Not Units.
At Innovation Day 2026 in Shanghai on October 29, AGIBOT introduced Worken, a proposed score for productive robot hours, alongside GO-3 Preview, its WRAM embodied foundation model, and the G2 Max heavy-duty robot. The company still sells the shipment story—trackers put it first in H1 2026—but Peng Zhihui, its co-founder and CTO, said the next question is how much useful work robots deliver autonomously. AGIBOT published the Worken formula and benchmark wins on RoboDojo, RoboTwin, and RoboColiseum, yet no customer has published a Worken total, and the output-efficiency term still needs task-specific human baselines.
AGIBOT, the Shanghai humanoid maker that independent trackers put at the top of H1 2026 shipment tables, held Innovation Day 2026 on October 29 and opened with a different scoreboard. Co-founder and chief technology officer Peng Zhihui introduced Worken, a proposed measure of "Effective Intelligent Work Hours" meant to track how much useful work robots perform on their own, not how many bodies leave the factory (AGIBOT release). The same stage carried GO-3 Preview, a WRAM embodied foundation model, the G2 Max industrial robot, and Link-U OS 1.0, plus a list of factory and public-service deployments from Longcheer tablet lines to Guangzhou Metro.
Our view: the company winning the unit-count race is now arguing that unit counts are the wrong metric. That is a coherent pivot for a vendor moving robots from demos into factories and theme parks, and it is also convenient timing. Shipment tables still dominate investor decks—Roboterminal's H1 piece showed two trackers agreeing on AGIBOT first and Unitree second (shipments report). Worken does not replace those tables until someone publishes autonomous hours and task output against a human baseline. AGIBOT has the formula. It does not yet have audited Worken totals from a paying site.
What Worken actually multiplies
The release defines Worken as active robots times autonomous working hours times an output efficiency ratio (AGIBOT release). Each term narrows the count. "Active" excludes inventory. "Autonomous" excludes powered-on time spent waiting for a human reset. "Output efficiency" compares finished work to what a human would deliver on the same task under defined standards.
Peng Zhihui's quote on stage frames the shift: after progress in hardware, models, and manufacturing scale, "we need to answer a more practical question: how much productive work are they actually delivering?" (AGIBOT release). The company calls Worken evolving and says it will refine benchmarks through operating data. That honesty matters. Without published baselines per task—totes per hour, inspections per shift, successful shelf scans per day—the third multiplier is whatever AGIBOT and a customer agree to call 100 percent.
A joint industry report from China Economic Information Service and the Shanghai Industrial Innovation Center of CAICT also went out at the event; AGIBOT says it includes Worken and deployment examples (AGIBOT release). The report is a context document, not a customer scorecard.
GO-3 Preview and the learning stack behind it
GO-3 Preview is AGIBOT's next embodied foundation model, built on WRAM—World Reasoning Action Model—linking reasoning, outcome prediction, and action (AGIBOT release). The company demoed zero-shot and few-shot behavior and described "Physical Intelligence Scaling" across pre-training, post-training, and continued improvement after deployment.
Benchmark claims at the event: first in task success rate and process score on RoboDojo, highest overall on RoboTwin, first overall on RoboColiseum (AGIBOT release). Leaderboards are useful for research comparisons. They are not Worken. A model can top a benchmark while a store robot still fails on the fifth shelf.
AGIBOT's Chinese recap adds engineering detail: Genie Evolver, the post-training stack, runs on roughly one hundred physical robots plus simulation and world models (Chinese recap). GO-3 Preview is slated to open source by the end of 2026 (Chinese recap). Link-U OS 1.0—灵渠OS in Chinese—has the same end-of-2026 open-source target and is being donated to the OpenAtom Foundation (AGIBOT release, Chinese recap). Open weights and an OS do not prove autonomous hours. They lower the cost of trying.
Secondary reporting on GO-3 describes a convenience-store G2 trial where workflow success rose from 28.6 percent on day one to 85.7 percent after five days, with the robot adjusting after its own failures (Interesting Engineering). Treat that as a reported demo curve, not a customer KPI, until AGIBOT or the retailer publishes the task definition and intervention log.
Hardware and deployments named on stage
AHA3 is AGIBOT's third-generation humanoid hardware platform, integrating locomotion, interaction, and task hardware (AGIBOT release). The Chinese recap highlights joint liquid cooling, IP67-rated joints for harsh environments, and a dexterous-hand line from OmniHand 3 Lite through OmniHand 3 Ultra (Chinese recap).
G2 Max, launched at the event, is a heavy industrial robot with 21 degrees of freedom, dual-arm handling up to 40 kilograms, and a working height from 0.1 to 2.5 meters in the English release (AGIBOT release). The Chinese recap adds a 20-kilogram single-arm rating and 50-kilogram peak load (Chinese recap). Deployments cited include Longcheer tablet production, semiconductor packaging and test, CNC machine tending, logistics, Guangzhou Metro visitor service, and the Chimelong tourism complex (AGIBOT release). The Chinese recap quantifies some public deployments: about 120 G2 units at Guangzhou Metro, roughly 500 robots operating at Chimelong with a target of 1,000 by year-end (Chinese recap). Robot counts are back. Worken would ask what fraction of those hours were autonomous and how much acceptable work each hour produced.
Why shipment leadership makes the pitch louder
AGIBOT's Innovation Day narrative assumes robots are already moving into daily operations. That matches the shipment story investors already know. It also raises the stakes for a new metric. If the leader in units cannot show superior Worken per robot, the formula becomes marketing cover for volume.
The ecosystem numbers AGIBOT quoted—255 universities, 511 enterprises, more than 80,000 developers—describe reach, not yield (AGIBOT release). Yield is the argument Worken is trying to win.
Objections
"Any mature industry eventually measures utilization, not shipments." Fair. Warehouses track pick rates. AGIBOT is early to say the quiet part out loud. The missing piece is standardized baselines outsiders can audit.
"Benchmark wins prove GO-3 will raise output efficiency." Not automatically. Benchmarks test model skill under benchmark rules. Worken multiplies real autonomous hours on real tasks. The link has to be demonstrated site by site.
"Theme-park and metro robots are not factory ROI." Correct—and AGIBOT's release treats service and factory work as different output types (AGIBOT release). Worken needs separate human references for each, or the ratio collapses into storytelling.
Three things to check
- First customer-published Worken total. A logistics operator or manufacturer stating active robots, autonomous hours, and output efficiency for a named task would turn the formula into a comparable KPI.
- GO-3 open source drop. Watch for weights, training notes, and whether post-training tools like Genie Evolver ship with the model at year-end 2026.
- Chimelong at 1,000 robots. AGIBOT's Chinese recap targets roughly 1,000 robots at Chimelong by year-end (Chinese recap). If the park hits that count, ask for autonomous operating hours per role—not just opening-day choreography.