45. 3 Hours Inside a Humanoid Robot “Working” in a Factory | A Conversation with 智元机器人

Summary

• 智元 has designated 2025 as the “commercialization year” for humanoid robots: mass production in 2024, commercial deployment in 2025. Held on the factory floor at 富临精工, the livestream was framed by 卫诗婕 as “the world’s first launch event bold enough to use a livestream to show a robot carrying out regular work in an industrial setting.” The first batch of 4 A2W robots worked continuously on-site, with cycle time optimized to roughly 40 seconds per tote. The robots can operate 24 hours a day and use battery swapping, but cycle time and success rates still had room to improve as of the livestream.
• 富临精工’s 邓扬 put a stark cost case on the table in person, making it the most investment-relevant demand checklist in the episode. A human moves one tote in roughly 30–35 seconds versus about 40 seconds for the robot; the robot sells for “several hundred thousand yuan,” before energy costs and the on-site service team. The payload has risen from 10kg to 15kg, while the dual-arm design is rated for 25kg. “At 20kg, the workshop could cover roughly 80% of its totes; at 30kg, it could essentially take on the whole job.” 邓扬 sees a 50% decline in total cost over the next 3 years as possible; 王闯 called that “far too conservative.”
• 智元 is taking an intensely pragmatic approach to the technology stack: it is explicitly not deploying VLA directly today, instead using “small models + rule-based control + data training.” Industrial success rates must reach 99.9%: “If you’re handling 800 totes a day and can’t get to three nines, the factory won’t let you in.” VLA still falls short on both success rate and cycle time. 王闯 compared the path with autonomous driving’s evolution from CNN to Transformer, Occupancy, end-to-end, VLM and VLA; the company’s end-to-end model is GO-1, released in March 2025.
• The industry’s division of labor is taking shape as a three-tier pyramid—robot platform makers, deployment integrators and manufacturing customers—with the integration layer seen as fertile ground for multiple listed companies. 安诺智能 was founded in February 2025, with 智元, 富临精工 and 巨星新材料 among its shareholders. Using the collaborative-robot era’s 克来机电 and 江苏北人 as comparables, 杨曾 said there could be dozens or even hundreds of mid-layer system integrators serving hundreds of thousands of industrial producers.
• The data flywheel was the episode’s recurring theme, with real-robot failure data explicitly defined as a moat. 王闯 cited a customer whose black interior caused recognition to fail: “That data from a real-world scene is extremely valuable; other robots may never have encountered it, and may not even have imagined it in simulation.” As digital-world data approaches exhaustion and Scaling Law slows, multimodal real-robot data from factories becomes a critical resource. But 邓扬 stressed that process data often contains a manufacturer’s core parameters and competitive edge, so it cannot simply be collected or opened up.
• “Hesitation” was defined as the dividing line for generality: during an on-site generalization test, the robot paused after a tote was deliberately moved and misaligned, analyzed the situation, and closed the loop autonomously. 卫诗婕 argued that A2W “hesitating for a moment” appears to reduce efficiency but actually reflects greater intelligence. 王闯 summarized it this way: “That thinking process in the middle is precisely a marker of embodied intelligence…the real dividing line, the thing that determines whether it is general-purpose, is that thinking.” A compound robotic arm moves faster precisely because it does not perform this kind of closed-loop reasoning; a small environmental change can trigger an alarm.
• The US-China comparison produced a counterintuitive conclusion: demand is more urgent in the US—labor is expensive and policymakers want manufacturing back—but many robotics companies cannot escape the lab, with the key constraint being access to real deployment resources and engineering capability. 王闯 sees no absolute algorithmic moat; data is often what matters. China’s complete industrial chain, open customers and ability to engineer down costs are advantages. A lab can fail 99 times and film the one success; a factory must perform 1,000 times and fail only once. Overseas hydraulic systems can cost more than $1M, making them difficult to design into a machine that actually works on the floor.
• The methodology for selecting use cases was itself a lesson: 智元 paid tuition on two “cool” applications—loading and unloading powder-coating lines with a 2mm tolerance, and loading and unloading anodizing lines with inconsistent hand-welded racks—before returning to the seemingly simple task of tote transfer. 邓扬 proposed a multidimensional “ladder chart” covering navigation and localization, arm closed-loop accuracy, payload and other variables to rank every factory use case systematically, while frontline employees would be asked to submit ideas for prizes. The host’s formulation was exact: the customer does not need “a sports car delivering takeout”; it needs the robot doing work suited to it and capable of generating ROI.

Deep dive

1. Setting the Scene: Three Sides Together for the World’s First Livestream of Regular Factory Operations

• The event took place on the production floor at 富临精工 in Mianyang, Sichuan, with the line still running. The A2W at center stage is a wheeled, humanoid-style industrial robot: 智元 developed the platform, 安诺智能 handled deployment, and 富临精工 put it into production. 卫诗婕 called it “the world’s first launch event bold enough to use a livestream to show a robot carrying out regular work in an industrial setting.”
• The 3 guests represented the 3 layers of the industry: 王闯, head of 智元’s general-purpose product line; 邓扬, from 富临’s engineering technology center, responsible for bringing in new processes and equipment; and 杨曾, chief engineering algorithm officer at 安诺智能.

2. 王闯’s Path: From Lithography Machines and DJI to Robot Vacuums—and the “Unreachable” Humanoid

• 王闯 completed his PhD researching lithography machines, joined DJI after graduation to work on LiDAR and autonomous driving, and took the technology through automotive-grade mass production. He later founded a robot-vacuum company. Using drones and robot vacuums as examples, he said both categories eventually reached mass production at the million-unit scale, with him taking products from zero to one twice.
• His candid admission before joining 智元 in 2024 is worth recording: “My understanding of humanoid robots was not deep enough…the image in my head was Honda’s ASIMO.” At the time, he thought it might take at least 6 months to 1 year of study to get up to speed. After entering the field, he said he was “refreshing his understanding almost month by month, day by day.”

3. Why 2025 Broke Out: The Spring Festival Gala and the Shift from Mass Production to Commercial Deployment

• 王闯 offered 2 causal chains. The Spring Festival Gala “gave many people who had never seen a robot their first look at one,” providing basic public education. More important, leading companies and industry partners “achieved mass production in 2024 and commercial deployment in 2025.”
• 远征A1, represented on-site by the assistant host “晶晶,” is already working at multiple customers. A2W shares substantial technology with A1 and A2. 智元’s product team has roughly 200 people, with the underlying software, hardware and part of the supply chain shared across products. A2W is designed for the dull, repetitive and physically punishing work that people do not want to perform in industrial settings.

4. Why Mass Production Is Hard: Supply Chains, Degrees of Freedom and Consistency

• The first mountain was the supply chain. Last year, the humanoid-robot ecosystem was “still extremely immature.” Joints and arm components for humanoids face different requirements from their conventional counterparts, particularly around torque density, weight reduction and noise. 智元 worked with upstream and downstream partners to develop components that were cheaper, more reliable and suited to humanoid robots.
• The second mountain was complexity. 远征A2 has 40 degrees of freedom, while A2W has 22 active degrees of freedom. More degrees of freedom expand the range of possible tasks, but also raise the bar for manufacturing complexity, process inspection and consistency control. Small errors in the upper limbs can be amplified at the end effector.
• The final requirement was to give every robot nearly identical performance “with one codebase and one model.” During last year’s year-end mass-production push, roughly 100 people—including 王闯—lived at the factory, while testing consumed more than 100 robots cumulatively. CCTV reporters once arrived at 智元’s production line to film without an appointment.

5. 富临’s In-House Electric Joint: Bringing Automotive Cost Discipline Upstream

• 邓扬 said 富临 began developing electric joints in 2023 and quickly encountered 3 challenges: high load, meaning “more power while becoming lighter”; control precision, including encoder design and input-output closed-loop control; and long life—“you can’t buy it today and have it fail after a week of work.” 富临 manufactures key joint components in-house, with material selection and machining precision both affecting durability.
• 王闯 added that the joints must be durable, precise and powerful—and “cheap.” 邓扬’s confidence comes from the automotive industry: 富临 has experience in large-scale production, manufacturing and lean operations, while the auto market imposes severe cost pressure. That gives it confidence in transferring automotive component-manufacturing know-how to electric joints.

6. Autonomous-Driving Talent Spillover: 杨曾’s Pivot and the Industry Trend

• 杨曾 joined US-owned Tier 1 supplier Harman in 2011 and helped develop the company’s first road-testable L3 prototype. He later worked at 上汽-backed 创时智驾, where he was primarily responsible for mass-producing intelligent-driving domain controllers. In his view, a smart car is also a robot, but one with only 4 wheels operating on the road and no arms, limiting its ability to affect the physical world.
• He confirmed that the talent shift is structural. Autonomous driving has developed for more than 10 years; L2++ is beginning to scale commercially, while L3 is on the eve of mass production. Combined with a weak auto market, that has created a spillover of autonomous-driving talent into the hotter embodied-intelligence sector.

7. What Is Embodied Intelligence—and Does A2W Count as Humanoid?

• Referencing Nvidia’s 黄仁勋, 杨曾 said embodied intelligence comes in 4 forms: autonomous vehicles, drones, robot dogs and humanoid robots. All are “AI plus a body”; the body simply changes with the application.
• A2W replaces legs with wheels because factory floors are flat, making wheeled robots quieter and more efficient. Its upper body is similar to A2’s, with 2 seven-degree-of-freedom arms, and it can use both grippers and dexterous hands. 卫诗婕 concluded that the real humanoid barrier may be the ability to think, distinguish and act like a person—not merely having 2 legs.

8. Why Build Bipedal Robots at All? Environmental Compatibility and “Being of the Same Kind”

• 王闯 is not doctrinaire: “We are not saying there is one form we absolutely have to build.” The team tried using a bipedal robot to move totes but found it less efficient, more prone to vertical swaying during transport and noisier. For factories, wheels made more sense.
• Bipedal robots have value elsewhere. Tables, chairs, shelves and other facilities in the human environment “are all designed for humans,” allowing a humanoid robot to reach them naturally without modifying the physical environment. A humanoid appearance and more advanced brain may also create a sense of familiarity.
• The ultimate vision is a general-purpose robot: different environments should not require a new body or hardware redesign. In some cases, a different model—or even the same model—should be enough to handle different jobs.

9. From Standing Unsteadily to Running: Reinforcement Learning as the Key to Humanlike Gait

• 王闯 recalled last year’s WAIC rehearsal. Organizers only required robots to walk for more than 10 minutes; 远征A2 used an early development prototype and walked for 13 minutes, making it one of the longest-running robots that day, though it still suffered issues such as overheated joints.
• Compared with this year, he sees a clear shift from bent-knee walking toward a more humanlike gait, driven partly by reinforcement learning. Model-predictive control previously delivered high precision but struggled to adapt to different surfaces or follow trajectories like a person. Those capabilities have now improved.
• Commercial anecdotes included several robots drumming together at a customer’s opening and a robot writing more than 130 calligraphy pieces in 1 day at the MWC China Communications Summit. 王闯 sees interaction intelligence as one of the most direct benefits of large models, making showroom explanations and interactions more natural.

10. “Some People Believe Only After They See; Others See Because They Believe”

• This belief narrative ran through the entire event: investors backed 智元 at the prototype stage, while customers opened their facilities for co-development before the machines were mature. 王闯 said, “We believed first, then made everyone see it.” They were a group that believed and worked to make others see.
• 卫诗婕 described the livestream as a way to witness a frontier technology at a very early stage of commercialization. The process requires believers across the industrial chain to watch it unfold together.

11. The Customer’s Field Report: Cycle-Time Gap, the 24-Hour Equation and Autonomous Recovery

• 邓扬 gave an unsparing assessment: “Cycle time has not yet reached our standard.” The robot takes roughly 40 seconds per tote versus 30–35 seconds for a human. But it can work 24 hours a day and uses battery swapping; “once you factor in meal breaks, it is about the same as a person.”
• 2 things stood out to him. On perception and obstacle avoidance, the robot says, “Please move aside—you’re blocking my work.” On abnormal-condition recovery, it can still grab a tote after it has been placed askew, keeping the line running continuously.
• Cost pressure came up repeatedly. A factory must pursue economic efficiency, so the equation includes not only the robot’s purchase price but also energy consumption and the cost of the on-site service team.

12. Payload and Use-Case Expansion: 20kg Covers 80%, 30kg “Covers the Lot”

• The key payload progression is straightforward: totes initially weighed roughly 10kg and now reach about 15kg, while the dual-arm design capacity is 25kg in total. 邓扬 estimates that 20kg across both arms would cover roughly 80% of the workshop’s totes; at 30kg, the robot could handle essentially all related work.
• 邓扬 also envisions robots working directly at the line with AMR carts, sending totes to storage racks. Warehouse picking, palletizing, depalletizing, line-side feeding and empty-tote collection could all eventually be explored. Technically, he considers these directions relatively accessible.

13. 王闯’s Response: The 160-Second-to-40-Second Slope and the Discipline of Owning One Use Case

• When the workflow first came together several months ago, it took 160 seconds per tote. After months of work, that fell to roughly 45 seconds and then to 40 seconds on-site. 王闯 believes reaching 30 seconds is entirely possible.
• Typical factory use cases fall into 3 categories: tote loading and unloading, sorting and assembly. The team chose totes because they were the closest match to current technology and customer demand. Industrial applications place an especially high premium on cycle time and success rate, so generalization must be constrained rather than pursued everywhere at once. The plan is to scale tote handling first; only with scale can costs come down.

14. The On-Site Generalization Test: The Robot’s “Hesitation” Is Intelligence

• The test deliberately moved a tote to the wrong position. After arriving and finding nothing where expected, the robot began analyzing, located the displaced tote, used its QR code to confirm that it contained the same material, and continued delivering it to the designated position, completing the loop autonomously. The team then placed the tote at an angle to test bimanual coordination and force control.
• The event’s key reversal was articulated by 卫诗婕 and confirmed by 王闯: a compound robotic arm moves quickly, making it easy to assume it is more intelligent. A2W “hesitated for a moment,” appearing less efficient but actually demonstrating greater intelligence. 王闯 summarized the dividing line: “What determines whether it is general-purpose is that thinking.”
• The robot also shows a humanlike preference for efficient positioning. When moving the tote on the opposite side, it decides that “it is easier to move it from this side” rather than reaching awkwardly from a distance. It can reach, but chooses the better stance.

15. The Fundamental Difference from Compound Robotic Arms: Closed Loops, Registration and a 20-Minute Line Change

• Compound robotic arms execute fixed programming at high speed and repeat it. A small change in environment or lighting can stop them, while an offset pallet position may trigger an alarm. A2W is positioned as a general-purpose robot for flexible manufacturing: before grabbing, it reads the information on the tote’s label and determines both the material and destination.
• The deployment tool is called “registration”: set where material A and material B should go. Behind the robot is a 3D map; storage racks can move, and the robot analyzes their approximate locations and closes the loop on its own. Moving to another workshop takes 10–20 minutes to rebuild the map and reissue tasks.
• 王闯 defines embodied intelligence as a human partner rather than a simple tool. That partner is “still a child” for now, with limited capabilities; when it encounters an exception it cannot handle, it calls a human remotely.

16. 2 Abandoned Use Cases: Tuition for Chasing the Coolest Applications

• Early on, the team was ambitious and believed “the harder the scene, the more we should challenge it.” The first lesson was loading and unloading powder-coating lines, involving more than 1,000 types of architectural hardware, including hinges, latches, lock cylinders and handles. The most difficult parts had screw holes around 5mm wide that had to fit hooks roughly 3mm thick, leaving only 2mm of tolerance. Low standardization across parts and hooks made many hole-based components impossible to handle.
• The second lesson was loading and unloading anodizing lines. Metal rings for smartphone cameras had to be placed on spring hooks and then fitted into racks with more than 10 layers. The racks were hand-welded and inconsistent, and they deformed with use; the tightness of each clasp also varied. Processes requiring 1–2mm precision were difficult, while placing parts onto machine tools with roughly 3–4mm of tolerance had a high success rate.
• The team’s review identified 2 causes: its capabilities were insufficient at the time, and the tasks demanded a high level of tactile sensing, force control and overall perception. The timing was also wrong for commercial deployment.

17. The Pivot to Totes: Customer First—“It Would Be a Waste Not to Have Something This Big Move Bricks”

• A researcher asked, “Why not do tote transfer?” The team found a genuine customer pain point, and some customers approached it proactively. After seeing A2W, one customer said, “It would be a waste not to have something this big move bricks.”
• Totes may look broadly similar, but they come with lids, open tops and bagged parts, while size and weight also vary. 王闯 believes that collecting varied data can enable generalization and that tote handling is technically a better first step than the previous 2 use cases.
• Rather than persist with scenes that were hard to commercialize, the team followed the principles of “customer first, product above all” and changed direction. Lab development took more than 3 months to bring cycle time down from more than 160 seconds. About 1 month ago, the team judged the robot ready to “work” at a customer site, deployed 4 units and stationed the R&D team there for 1 month.

18. Why Livestream It? Openness and Confidence—“Like a 3-Year-Old Who Just Learned to Ride a Bike”

• 王闯 stated the livestream’s purpose plainly: “Everyone has questions. They ask what embodied intelligence can actually do and when it will really work, instead of seeing Demo videos every day. We wanted to answer that question.” The tote use case may look simple, but it tests the full technology stack and platform stability, while offering real customer value and a path to future generalization.
• The team was equally clear about its limits: “We wouldn’t dare say it is already perfect.” Cycle time and success rate are not final. 王闯 compared the robot to a 3-year-old who has just learned to ride a bicycle; greater challenges lie ahead, and the process will require patience and time.
• As of the livestream, there had been no errors on-site. 王闯 added that the current interference may simply not have been strong enough; longer operating hours will likely reveal new problems.

19. Who Is 安诺? Shareholding and the Integrator Pyramid

• 杨曾 described 安诺 as “a leading supplier for the engineering deployment and mass-production commercialization of humanoid robots.” Founded in February 2025, the company was only about 6 months old, with 富临 as its first deployed scene. Shareholders include 智元, 富临精工 and 巨星新材料. The latter supplies magnetic materials and is an upstream supplier of core motor materials; 杨曾 called it one of China’s largest permanent-magnet producers.
• The division of labor is clear. 智元 handles the platform, core underlying algorithms and the first benchmark workflows. 安诺 uses that base to deploy the technology at more customers and in more scenarios.
• Drawing on the collaborative-robot precedent, 杨曾 cited system integrators such as 克来机电 and 江苏北人, noting that system integration accounts for 70% of some companies’ business. His pyramid has a few platform suppliers at the top, dozens or hundreds of system integrators in the middle, and hundreds of thousands of industrial producers below. The middle layer could produce a significant number of listed companies.
• When the host asked the price on-site, the answer was “several hundred thousand yuan.” With 安诺 only recently founded and 富临 its first deployment, 卫诗婕 called it a gathering of “3 people willing to be the first to eat the crab.”

20. The 7-Step Deployment Process: How Hard Is Commercialization, Really?

• 杨曾 broke the process into 7 steps: build a map with LiDAR; register totes and loading ports visually and collect a small amount of data to optimize the model; adapt gripper width, grasping position and form; connect the business workflow using atomic capabilities supplied by 智元; connect communications with the production line; continuously optimize cycle time, the visual model and action sequence; and finally run stress tests under real operating conditions.
• Communications include notifying logistics to remove an empty pallet and bring a full one after a tote is moved, as well as using QR codes to identify which loading port should receive each tote.
• A concrete counterexample showed why parallelization is not always beneficial: after placing a tote, the robot may retreat while retracting its arm, and the arm can sometimes hit the rack. The latest effort took more than 3 months of lab development and 1 month of concentrated on-site work. Future deployments will feed the experience back into the model, steadily improving efficiency for the next customer and next scenario.
• 卫诗婕 compared deployment to 杨利伟’s first crewed space launch, arguing that experience will gradually lower both deployment difficulty and cost.

21. Why US Robots Struggle to Leave the Lab: The Scarce Resource Is Deployment, Not Just Algorithms

• 王闯 believes the US has stronger motivation to build robots because labor is more expensive and there is a greater push to bring manufacturing home. But when US robotics companies try to deploy, they need customers with large-scale production, advanced lines and a willingness to open up for co-development. That combination is scarce.
• China has a complete industrial chain, access to suitable customers and the ability to build platforms that are relatively stable and inexpensive. 王闯 sees no absolute algorithmic moat; in many cases, data is the key. There is also a large pool of Chinese engineers and rapid engineering iteration.
• The cost and standards gap is stark. Overseas robots can cost several hundred thousand dollars, while hydraulic systems can reach more than $1M, making them difficult to design for real work. A lab can fail 99 times and film 1 success; a factory must perform 1,000 times and fail only once. 邓扬 added that all 3 sides need to stay pragmatic, start with the simplest actions and give the new technology time.

22. Data Flywheel I: 4 Data Categories and “Failure Data Is the Moat”

• 王闯 divided collection into 4 steps: recognition, learning how to grasp thousands of tote types and how much force to apply; operation, or the “cerebellum,” learning which grasping methods maximize efficiency and success rate, along with full-body coordination such as bending for a low tote or leaning forward for a distant one; movement, collecting large volumes of map and environmental data; and placement, covering different destinations such as flow racks, conveyors and shelves.
• One movement-data example came from a customer whose black interior caused recognition to fail. 王闯 said this kind of real-world data is highly valuable: other robots may never have encountered it, and even simulation may not have anticipated it. Failure data feeds back into algorithm improvements and becomes a capability moat.
• If the robot cannot reach a shelf even at full extension, that failure can inform the next generation’s degrees of freedom and workspace design. The end state is a robot that sees a new tote and can say, “I’ve seen this one before; you don’t need to teach me again.”

23. Data Flywheel II: Physical-World Data Takes Over from Scaling Law, but Process Data Is Naturally Confidential

• 卫诗婕 connected the dots: one reason large-language-model Scaling Law is slowing is that digital-world data is nearing exhaustion. The next wave of data mining will move from online digital environments into the physical world, where deployment collects multimodal physical data.
• 王闯 explained why real-machine data matters: machines working in the real world obey physical laws, and enough accumulated data can support inference of a world model for industrial sites. 邓扬 confirmed that much process data consists of core parameters developed through years of iteration and represents a manufacturer’s competitive edge; it cannot simply be opened to outside collection.
• The contrast with autonomous-driving perception is sharp. A self-driving system may only need to know that “there is a person 100 meters away.” A robot must identify a tote a few dozen centimeters away, determine where the end effector should land, where to grip most securely and how the other arm should coordinate. Spatial perception and precision requirements are much higher. 王闯 classifies full-body coordination and planning as the “cerebellum.”

24. OTA, Customer Self-Training and the Question: “Then What Does 安诺 Do?”

• 智元’s OTA pipeline is already live. A cloud platform shows the software version on every machine, and the company pushes an average of 1 update to customers every 2–3 months. It is also considering opening training accounts to important customers with the capability to use them.
• 邓扬 said 富临 is considering building a team to train visual models. 王闯 believes that if the data remains inside 富临’s factory, 富临’s deeper knowledge of the site will accelerate generalization. Customers solving generalization themselves would represent another step forward for the industry.
• The customer’s demand is explicit: future generalization deployments cannot keep adding costs beyond the hardware. Customers need to know whether they will pay more, whom they will pay and how much.
• When 卫诗婕 asked what 安诺 would do if 富临 handled everything itself, 杨曾 said that using data to redefine the upstream “brain” is 安诺’s medium- to long-term direction. A small number of customers may eventually deploy independently, but most will still need integrators, while 安诺 retains value in training from zero to one for new scenes and new totes.

25. The Technology Stack: Why It Is Not VLA Yet

• 王闯 sees VLA’s advantage in cross-scenario generalization and zero-shot capability. Faced with a pile of supermarket snacks, for example, a VLA system might not need to have seen each item before picking it up. Its current weakness is that it cannot yet push individual industrial capabilities to a high enough level; both success rate and cycle time remain insufficient.
• An industrial robot may need to handle 800 totes in 1 day, requiring a 99.9% success rate or the factory will not allow it inside. A2W currently uses “small models + rule-based control + data training”: recognition starts with a small model and gradually builds an expert model for totes; grasping uses rules first to protect precision and success rate; movement and obstacle avoidance are continuously optimized; placement combines shelf positions with autonomous analysis.
• 王闯 compared the route with autonomous driving’s evolution from CNN to Transformer, Occupancy, end-to-end, VLM and then VLA. 智元 is researching and positioning around VLA, but does not automatically choose the most advanced technology for deployment. It chooses what is most efficient for the customer. The company’s end-to-end model is GO-1, released in March 2025.

26. 智元’s Full-Stack Thesis: 1 Platform Plus 3 Forms of Intelligence

• 王闯 summarized 智元’s point of differentiation: “From the start, we deliberately chose a full-stack technology strategy”—1 platform plus 3 forms of intelligence. Motion intelligence is the foundation; interaction intelligence and task intelligence create customer value.
• The product matrix includes the 灵犀 line for small humanoid robots, the Genie精灵 line for data-driven end-to-end systems targeting commercial services, and the general-purpose line, including the full-size 远征A2 humanoid and A2W industrial robot. Shipments and platform stability for products such as A2 and A2W are relatively advanced in the industry, while subsequent products such as G1 are also being developed.
• Motion intelligence is not only about walking but also upper- and lower-body coordination. Interaction intelligence has begun large-scale deployment in real customer settings, while A2W is a flagship example of task intelligence entering the field. Getting customers to use the product matters more than making something flashy in the lab.

27. The Engineering of Cycle Time: How 160 Seconds Became 40

• Cycle-time improvements came from faster recognition, parallelized actions and trajectory optimization. The model had to be compressed for deployment; obstacle recognition went from requiring roughly 0.5 seconds of confirmation to about 0.1 seconds. The robot scans its surroundings 10 times per second to maintain safety.
• The robot now walks faster and turns while raising its arm, both products of whole-body planning and control. Its trajectory also improved from straight-line movement to smoother routes selected for each point.
• The most interesting engineering detail is that the robot treats its own arms and the object it is carrying as obstacles. Autonomous vehicles never face this problem because their sensors are fixed and they have no limbs that continuously change during work. Data from 4 tote types, along with simulated changes in lighting, color, position, shape and stacking patterns, all fed back into the deployment.

28. How to Rank Use Cases: Ladder Charts and Frontline Inspiration

• 邓扬 wants a replicable evaluation method that works like a quality-audit checklist: determine whether a use case is constrained by payload, oil, water and other physical conditions, or by navigation precision, and use the answer to tell 智元 and 安诺 where the next-generation product needs to improve.
• 王闯 proposed a “use-case difficulty ladder chart,” modeled on CPU ranking charts but built around multiple dimensions. Navigation and localization require roughly ±5cm accuracy, while autonomous closed-loop arm control requires about 5mm. Use cases demanding precision within 5mm can wait until this year’s capabilities improve. With current design payload above 20kg, 30kg scenarios can also be excluded for now.
• The final test is customer value: can the robot do the work well and quickly, and is the ROI high enough? 卫诗婕 suggested asking 富临’s frontline employees to submit use cases for prizes. 王闯 said 智元 could provide the awards.

29. The Answer to Job Anxiety and the Smart-Factory Vision

• 邓扬 responded directly to comments about jobs. 富临’s various sites employ more than 5,000 people, and the goal is not simply to replace workers with robots. Robots should first take on heavy physical labor, while people manage equipment and eventually manage robots. These roles require accumulated knowledge and experience, and compensation may rise over time.
• His vision is a smart factory built around interaction among people, robots and equipment: robots interact with machines, while people interact with robots.
• Procurement rules also draw a clear line against hype: “If an industrial robot can do the job, I absolutely will not use a humanoid robot, because industrial robots are already highly mature.” 卫诗婕’s analogy was that the customer does not need “a sports car delivering takeout”; it needs the sports car doing work suited to it.

30. The 1-to-3-Year Outlook: Model Breakthroughs Take Time, but a 50% Cost Cut May Still Be Conservative

• 杨曾’s forecast was measured: “A breakthrough in models may still take some time.” 安诺’s priority is to expand from the existing template into more scenarios and customers, collect data and wait for the models to break through.
• From the electric-joint supplier’s perspective, 邓扬 believes total costs could fall 50% over the next 3 years. 王闯 responded on the spot: “That’s conservative—absolutely conservative.” 邓扬 believes falling costs will encourage more users to try the technology even if models and training remain imperfect; cost is the central condition for commercialization. He said the business unit and innovation engineering division generate more than RMB1B in revenue.
• 富临 was willing to be an early adopter because “we believed, and then we saw,” but not simply out of conviction. The decision was based on a concrete site assessment: jobs that conventional industrial robots can handle will not be handed to humanoids, while today’s tote-moving task was judged suitable for A2W.
• In the closing bonus, 晶晶 read out A2W’s results for the day: “There were no collisions or intrusions throughout the entire process.” When people entered the work area, the robot also stopped automatically.