After entering real-world scenarios, robots still face several hurdles

(Source: Robot Global Insights)

In front of the shelves of a simulated supermarket, humanoid robots identify products, pick them, scan them, and pack them according to orders; on the other side, in the quality inspection arena, they must unscrew bottle caps, place test tubes, and operate centrifuges, while materials and fault points will change randomly...

The just-concluded 2026 Global Developer Pioneer Conference and International Embodied Intelligence Skills Competition moved a series of real industrial challenges into the arena. The competition looks not only at whether the movements are smooth, but also whether the robots can still complete the tasks when the environment changes.

  The 'stress tests' on the arena reflect a shift underway in the embodied intelligence industry. Recently, Xinhua Economic Information Agency and the Shanghai Industrial Innovation Center of CAICT jointly released the '2026 Embodied Intelligence Industry Innovation Report' in Shanghai. The report states that in 2026, embodied intelligence will begin transitioning from development demonstrations to routine operation in real-world scenarios, entering a deployment growth phase. The yardstick for judging industry progress is also shifting from how many products are launched or how many units go offline, to how many robots truly enter factories, warehouses, and stores, continuously complete tasks, and deliver results.

From Demonstrating Movements to Completing Tasks

  In recent years, humanoid robots have run faster and moved more dexterously, frequently attracting attention. However, a spectacular demonstration does not equal a replicable production solution. At the 26th China International Industry Fair held in Shanghai from October 12 to 16, questions in the robot exhibition area became noticeably more specific: what level of precision, what load capacity, how long deployment takes, and whether continuous operation is stable.

At the Industry Fair, AGIBOT took 'industrial deployment state' as its main thread, placing the Elf G2 series robots into four simulated workstations: optical module plugging/unplugging, cylinder head valve assembly, automotive seatbelt metal shell loading/unloading, and palletizing of material boxes. From sub-millimeter precision operations to 50 kg peak heavy-load handling, the exhibition stand showcased the robot's capability boundaries; to prove these capabilities can be scaled for replication, long-term operation data on production lines must also be examined.

AGIBOT Elf G2 robot performs a demonstration of “assembly and loading station” operations. Photo by Xinhua News Agency reporter Fang Zhe

  The report recorded a period of testing by AGIBOT at Longqi Technology the Nanchang factory: 8 Elf G2s accumulated over 64 hours of operation in a 6-day test, completing 64,828 operations, with a comprehensive operation success rate of 99.99%. This set of slightly dull data is, conversely, closer to the issues that industry truly cares about: whether robots can run stably, how much human intervention is required, and whether the comprehensive cost is worthwhile.

  The report summarizes industrial evolution along three 'XYZ' curves. The X curve from 2022 to 2025 was a development trial phase, focusing on solving whether robots can move stably and be mass-produced; the Y curve from 2026 to 2030 is a deployment growth phase, where robots must shift from completing movements to stably completing tasks; only afterward could it enter the Z curve characterized by general capability improvement and intelligent emergence.

  The term 'deployment state' does not mean moving a prototype to the production site, but rather that robots can operate routinely in real environments and continuously deliver results. Behind this lies a comprehensive set of system capabilities, including overall reliability, mass-production consistency, model generalization ability, data closed-loop, and O&M support. Zheng Zhongbin, Chief Expert at the Shanghai Industrial Innovation Center of CAICT, stated that demonstrations showcase optimal performance, while deployment requires meeting production line standards even under the most unfavorable conditions. The most urgent threshold currently lies in engineering reliability and acceptability, as well as forming a continuous evolution closed-loop through real-world deployment.

  Scenario selection has therefore become more pragmatic. Currently, embodied intelligence first enters task domains with clear boundaries and quantifiable results such as material handling, sorting, inspection, and loading/unloading, then gradually expands to complex environments like flexible assembly, commercial services, and households. The report calculates that if the success rate of a single action is 99%, the overall success rate for completing ten actions in sequence will drop to about 90%. For industrial production, this means robots must start with short-chain, high-certainty tasks and gradually extend the task chain during operation.

From Single-Point Skills to Learning by Doing

  As robots enter real-world scenarios, the difficulty lies not only in the mechanical body, but more in making three types of capabilities—movement, operation, and interaction—work in synergy.

  The report summarizes this goal as "the integration of three intelligences." Early robots often trained modules for movement, grasping, and dialogue separately; while individual capabilities could be impressive, information loss and error accumulation between modules would be amplified in complex tasks. The new technical approach attempts to use a unified model to connect perception, decision-making, and execution, enabling robots to both understand tasks and adjust actions according to environmental changes.

  According to the capability grading proposed in the report, current robot bodies and interaction capabilities are developing rapidly and are already in the transition from L3 to L4; motion intelligence is roughly at the L3 level; operation intelligence, which directly determines labor value, remains slightly below L3. Whether robots can stably complete tasks when facing unfamiliar objects, lighting changes, and personnel interference still requires more verification in real-world scenarios.

  Data represents another critical hurdle. Text and images can be obtained from the internet, but robot training data must record visual, force, tactile, and motion trajectory information, as well as understand why an operation succeeded or failed. Real-machine data incurs high costs, while simulated data has gaps from real environments. The report believes that embodied intelligence needs to shift from centrally collecting data in laboratories to robots learning while operating in commercial scenarios, forming a closed loop of deployment, data collection, training, updates, and redeployment.

  The recently concluded Global Developer Pioneer Conference further opened this iterative process to developers. The conference set up real-scenario training sessions and workshops where attendees could operate actual machines from companies such as AGIBOT and Unitree on-site; enterprise demands in retail supermarkets, quality inspection, and energy patrol were converted into competition problems and reusable "skill packages." For the industry, the value of the developer ecosystem lies not only in increasing the number of applications but also in enabling experiences from different robot bodies and scenarios to be verified, accumulated, and reused.

  This also explains why actual deployment volume is more closely watched than mere shipment figures. A robot operating for a long period serves as both a production tool and a data entry point. The more deployment scenarios there are, the richer the long-tail problems encountered by the model; improved model capabilities may open up more scenarios. However, whether this flywheel can spin depends on whether data formats can be unified, how privacy is protected, and whether data from different robot bodies can be reused. Zheng Zhongbin summarized these bottlenecks as "non-connected, non-trusted, non-exchanged": data standards are not connected, the credibility of simulated versus real-machine data lacks calibration, and mechanisms for data ownership and revenue distribution remain to be improved.

The inflection point of scaling must be judged by hard metrics.

  Market expectations for embodied intelligence are rising. The Development Research Center of the State Council's "China Development Report 2025" forecasts that China's embodied intelligence industry market size is expected to reach 400,000,000,000 yuan by 2030 and surpass one trillion yuan by 2035. The report also warns that sustainable industrial development still requires attention to generalization capabilities, high-quality data, edge-side computing power and power consumption, component durability, and system security—thresholds that must be crossed for large-scale deployment.

  Currently, exploration of scaled deployment is extending from factory workshops to more open cultural and tourism scenarios. On September 24, AGIBOT partnered with Chimelong Group to launch an embodied intelligence-themed park operation, with over 300 robots distributed across more than 100 interactive experience points, covering scenarios such as entertainment performances, science popularization and study tours, guided tours and shopping assistance, service retail, intelligent companionship, hotel services, and sports competitions. The company stated that this large-scale routine deployment is a pioneering exploration within the industry. For the industry, its observational value does not lie in whether robots "can perform," but in whether they can operate continuously amid fluctuating visitor flows, open interactions, and multi-scenario switching, and whether they can form maintainable and operable service samples.

Chimelong Spaceship Park opens with a routine "human-machine co-performance" grand opening dance. Photo provided by the interviewee.

  However, an increase in scenarios and deployment volume does not automatically mean the inflection point of scaling has arrived. The task boundaries, interaction safety, and operational rhythm of theme parks differ from industrial production lines, so their deployment experience cannot be simply replicated to manufacturing processes. The true threshold for scaling must fall on hard metrics such as continuous operation time, effective task completion volume, manual takeover rate, comprehensive costs, and system security. Precisely because of this, embodied intelligence still lacks a universally recognized value scale.

  Traditional industrial robots are typically evaluated by equipment price, uptime, or investment return rate per process. The report introduces that AGIBOT has proposed the concept of "effective intelligent working hours," attempting to measure robot output by combining the number of active robots, actual effective working duration, and task value density. This metric currently remains an enterprise exploration; whether it can become an industry-wide standard requires repeated verification in more scenarios.

  Safety is a harder constraint than cost. Robots directly act on the physical world; a single recognition error or loss of motion control could lead to real-world consequences. The report recommends integrating safety governance throughout hardware, algorithms, systems, and deployment processes, clarifying responsibility boundaries among developers, users, and regulators, while strengthening data minimization collection, desensitization, and encryption protection.

  Robots entering production and service scenarios should not be simply characterized as substitutes for humans. The report believes that, at this stage, the more visible path is human-machine collaboration, where robots undertake repetitive, high-intensity, or continuous tasks, release new job positions and opportunities, and activate productivity value. At the same time, employee rights, skills transformation, and safety responsibilities should be assessed to ensure coordination between technology application and human development.

  Moving from the exhibition stand to real-world scenarios means that embodied intelligence begins to undergo continuous testing in production and life. Lively action demonstrations will still attract eyes, but what will determine how far the industry can go will be another set of less eye-catching numbers: how long it can run continuously, how many effective tasks it completes, how much manual intervention is required, and whether the comprehensive cost can be accepted by customers. Zheng Zhongbin stated that “acceptable” is an important milestone for crossing the deployment threshold. (Sun Qing)

Massive information, precise analysis, all within the Sina Weibo APP