AgiBot releases Genie Studio 2.0: a one-stop embodied intelligence operation platform, enabling deployment of new scenarios in as fast as 7 days

AI WAVE | AgiBot Releases Genie Studio 2.0, Closing the "Last Mile" for Robots to Take Up Posts — Zhong Lilei, 2 hours ago. Deploy a new work scenario from 0 to 1 in as fast as 7 days.

On October 23, following up on the "first real-world scenario Physical Agent dataset," AgiBot released the second installment of its seven-part AI WAVE series — the one-stop, full-process embodied intelligence work platform Genie Studio 2.0. Centered on the AI WAVE series goal of "getting robots truly on the job," Genie Studio 2.0 closes the "last mile" for robots moving from model capabilities to real-world work, driving the transformation of robot capabilities into work solutions that can run sustainably and be replicated at scale.

As large models and embodied models keep setting new lab performance records, the industry faces a harsh reality: many robot demos are impressive, but once they enter a factory they suffer from "acclimation problems."

Getting a model to run is only the first step. To truly take up a post and do the work, a robot must understand on-site materials, interface with production line equipment, autonomously handle unexpected anomalies, and support dozens of devices running stably 24/7. Capability does not equal productivity; between the "demo state" and the "deployment state" lie the formidable barriers of development, orchestration, integration, operations and maintenance, and replication.

On October 23, as the second installment of AgiBot's seven-part "AI WAVE" series, the one-stop, full-process embodied intelligence work platform Genie Studio 2.0 was officially released. The platform consists of three core modules: the Flywheel embodied model iterative development platform, the Agent modular work orchestration platform, and the Monitor multi-robot fleet operations platform. Together they fully cover the entire lifecycle of "collection-training-testing-inference-orchestration-management," filling the key gaps that keep robots from moving from model capabilities to real on-the-job deployment, and enabling deployment validation of a brand-new work scenario from 0 to 1 in as fast as 7 days.

If Flywheel addresses "how a robot adapts to a specific skill," then Agent addresses "how different capabilities work together to complete a real work task," and Monitor addresses "how a robot runs stably after deployment, keeps improving, and is replicated at scale." Together, the three close the "last mile" for robots moving from model capabilities to real-world work. Stop One: Flywheel | The robot's "skills boot camp": quickly learning post-specific abilities

When general-purpose robots enter real factories, they face vastly different processes, materials, and precision requirements, and must undergo "on-the-job retraining." Flywheel is the core platform for completing this adaptation and iteration.

The complete closed loop consists of four steps: data collection → model training → simulation mock exam → real-machine deployment, with anomalous data flowing back to feed further iteration.

Step one: multi-source data collection. Flywheel uniformly manages real-machine, simulation, and open-source data. Real-machine data records the robot's visual information, motion trajectories, and execution states; for cases where collection costs are high, occurrences are infrequent, or testing risks are significant, simulation data can be generated via Genie Sim; the platform can also connect to open-source datasets such as AGIBOT World to expand task and scenario coverage.

Step two: efficient model training. With data in hand, customers can train image perception models, vision-language-action models (VLA), and real-machine reinforcement learning models on the platform, using tools such as the GO-2 embodied foundation model, model fine-tuning, experiment management, and compute scheduling. In real project validation, AgiBot's self-developed reinforcement learning training framework achieved hour-level convergence, with task success rates reaching 99%.

Step three: simulation testing. After model training is complete, the model must first pass a "mock exam." The Genie Sim Benchmark evaluation platform covers four dimensions and 42 task categories, supports onboarding in 5 minutes and completion of evaluation in 30 minutes, and keeps the error between simulation results and real-machine performance under 10%, intercepting most risks before real machines go live.

Step four: real-machine deployment + error-notebook closed loop. After robots are put into operation, failure and anomaly data return to Flywheel for the next round of training. Flywheel is thus also like a continuously updated "error notebook": every type of on-site problem solved accumulates reusable capability for subsequent deployments. Stop Two: Agent | The robot's "work orchestration system": assembling complete work flows like building blocks

Learning a single skill is still far from completing real business tasks.

Take machine tool loading and unloading as an example: a complete task includes dozens of steps such as receiving production instructions, navigating to pick up materials, identifying workpieces, grasping and transferring, interacting with equipment, and exception rollback. An error in any one link interrupts the entire flow.

Genie Studio Agent uses visual drag-and-drop orchestration to freely combine basic capabilities such as navigation, visual positioning, model inference, and logical judgment like building blocks, quickly building complete work workflows. The platform supports on-demand combined invocation and supports connecting model inference capabilities such as VLA and reinforcement learning. For complex scenarios and customization needs, the platform also provides a rich set of secondary development tools, allowing developers to connect custom business logic and communicate with third parties.

The orchestrated application can also be trial-run in a simulation environment first to check paths, collisions, poses, and process logic. Validated processes, skill nodes, and parameter configurations can also be packaged into reusable business templates—develop once, reuse across multiple projects. Station 3: Monitor | The robot fleet's "central command room": managing dozens to hundreds of robots to keep working continuously

A single robot can be debugged by an engineer on site; when dozens of robots are rolled out at scale, manual robot-by-robot operations and maintenance is completely unrealistic.

Monitor is a full-domain operations and maintenance platform for multi-robot fleets. It can centrally view each robot's device status, task progress, and operating data, and perform group management and task scheduling for robots.

When a robot experiences an anomaly, Monitor can record the fault, issue an alert, and support remote intervention and online debugging. Managers do not need to log in to each device and dig through logs one by one; instead, they can more quickly pinpoint the specific robot and task step.

The operating data, anomaly information, and handling results recorded by Monitor can also be returned to Agent and Flywheel for adjusting task flows and updating models. Going online is not the endpoint for a robot, but the starting point of continuous evolution. Real-world scorecard: rapid deployment in 7–14 days, with multiple benchmark scenarios completing large-scale validation

Currently, Genie Studio 2.0 has withstood real production-line tests in multiple real projects in manufacturing and rail transit.

Changzhou CNC machine tool project | Completing the journey from requirements to production-line deployment in 14 days

Centering on the loading and unloading needs of 6 CNC machine tools, robots and automated guided vehicles were introduced for collaborative operation. Through Genie Studio 2.0, the complete process from requirements analysis to production-line validation was completed in only 14 days. Flywheel completed data and model preparation, and Agent orchestrated navigation, material picking, machine tool loading and unloading, material placement, and exception handling into a complete flow. The robots first checked operating paths, action transitions, and potential collisions in the simulation environment before proceeding to real-machine validation.

During the on-site joint debugging phase, the robots connected with PLCs (programmable logic controllers), MES (manufacturing execution systems), vision systems, and tooling equipment, while Monitor synchronously connected device status and task information. After continuous testing, the relevant task flows, capability nodes, parameters, and interface configurations were organized into a project template, which similar production lines can quickly reuse.

JUNPO Intelligent BMS assembly line | 99.9% task success rate, 7×24-hour continuous production

At JUNPO Intelligent's pilot validation line, three Genie G2 robots work together to complete the loading and assembly of parts for automotive battery management system (BMS) controllers, finished-product boxing, and bin transfer. The two parties orchestrated tasks based on Genie Studio 2.0 and validated the flow in advance in the simulation environment. Facing highly reflective metal objects and placement accuracy requirements of less than 1 millimeter, the success rate of the relevant precision loading/unloading and assembly tasks stably reached over 99.9%, with a single operation cycle controlled within 13 seconds, and 7×24-hour stable production validation was completed.

Longcheer Technology tablet loading and unloading line | Unified operations and maintenance for a multi-robot fleet

At Longcheer's tablet computer loading and unloading line, the Genie Studio 2.0 platform manages multiple robots in a unified manner, tracking working time, operation counts, success rate, cycle time, and fault conditions. When a task anomaly occurs, operations personnel can see which node the process has executed to, and the relevant data also flows back to the cloud for problem analysis and subsequent optimization.

Guangzhou Metro service scenario | Large-scale deployment of 60+ devices

In the Guangzhou Metro station service scenario, Genie Studio 2.0 has supported the deployment of more than 60 devices. Through a global dashboard, the platform centrally displays indicators such as device status, task success rate, and mean time between failures across different stations; when a fault or warning occurs, the system can send message alerts and supports remote connections and quick commands, reducing reliance on on-site inspections.

Currently, embodied intelligence entering real work scenarios still needs to overcome multiple thresholds such as on-site debugging, system integration, long-term operations, and cross-scenario replication. As the second installment of "AI WAVE," Genie Studio 2.0 fills in not only the model development link, but connects data, models, robot skills, and on-site systems into a deployable, operable, and continuously optimizable work system.

As capabilities, processes, and templates continue to accumulate, robot deployment will gradually reduce its dependence on custom development from scratch, moving toward more efficient application delivery, more stable long-term operations, and lower overall cost of use. Enabling full-stack embodied AI technology to ultimately create value through real work scenarios is precisely AgiBot's goal in continuously advancing the implementation of work intelligence.

Source: Contributed article

Images in this article are from: 123RF licensed image library. Infringement report Return to sharing Related articles All comments