“接下来，是通用人工智能和具身智能发展的十年。”10月11日，极智嘉CTO李洪波在清华大学人工智能研究院学术年会暨京沪港AI学术交流会上发表主旨演讲，这家头顶“全球AMR机器人第一股”光环的行业龙头，被再次拉回聚光灯下。去年7月，极智嘉登陆资本市场，曾凭借全球近10万台商用机器人的落地规模，股价一度冲至近40港元/股，如今距高点已跌去八成。10月9日极智嘉发布最新公告，其全资子公司北京极智嘉拟出资5000万元入股北京深诣，交易完成后将持有后者20%股权。公告当日，极智嘉股价一度触及上市以来低点，最终以每股6.96港元报收；10月12日，极智嘉报收于6.8港元/股。
"The next decade is for artificial general intelligence and embodied intelligence development." On October 11, Geek+ CTO Li Hongbo delivered a keynote speech at the Tsinghua University Institute for AI Annual Academic Conference and Beijing-Shanghai-Hong Kong AI Academic Exchange. This industry leader with the halo of "the world's first AMR robot stock" was once again pulled into the spotlight. Last July, Geek+ landed on the capital markets, once seeing its stock price surge to nearly HK$40 per share on the strength of nearly 100,000 commercial robots deployed globally. Now it has fallen 80% from its peak. On October 9, Geek+ announced that its wholly-owned subsidiary Beijing Geek+ plans to invest 50 million yuan in Beijing Shenyi, acquiring a 20% stake after the transaction. On the announcement day, Geek+'s stock price once touched its lowest point since listing, finally closing at HK$6.96 per share; on October 12, Geek+ closed at HK$6.8 per share.
“如果你不把端到端这个技术做透，就很难在7×24小时、一年365天持续给客户稳定交付。”李洪波坦言，行业存在重演示、轻落地的痛点。从4米/秒漂移过弯的具身小脑，到应对百万级SKU的数据智能，公司试图用“先学再调”的解法穿越概念泡沫。无法否认的是，机器人在演示视频里的流畅画面，与真实场景中的稳定运行之间仍横着鸿沟。这家仓储机器人龙头的逆势加码，是穿越周期的超前布局，还是一场赌注？
"If you don't thoroughly master end-to-end technology, it is difficult to continuously deliver stably to customers 24/7, 365 days a year." Li Hongbo admitted the industry has a pain point of valuing demonstrations over deployment. From a 4 m/s drifting cornering embodied cerebellum to data intelligence handling millions of SKUs, the company attempts to cross the concept bubble with a "learn first, then tune" solution. It cannot be denied that there remains a gulf between the smooth images in robot demonstration videos and stable operation in real scenarios. Is this warehouse robot leader's counter-trend increase a forward-looking布局 through the cycle, or a gamble?
“不管是强人工智能还是强具身智能，范式转换的本质是自主感知、自主决策、自主操作。”李洪波用极智嘉自身的发展经历进行类比，称公司的起点是货到人模式，用机器人将货架搬到拣货员面前，帮助拣货人员将有效工作时间从30%提升至95%以上。不过，这尚属预设程序自动执行的弱人工智能与弱具身智能，真正的跨越发生在机器人可以开始处理极限工况。极智嘉专注于自主移动机器人的研发与应用，提供智能仓储物流解决方案。仓储机器人行业过去十年的增长逻辑是用自动化替代重复劳动，其技术前提是场景的结构化程度足够高、任务的可预测性足够强。而下一个十年的逻辑，将变成用智能体解决非结构化问题，这就不仅仅是技术参数的提升。
"Whether strong AI or strong embodied intelligence, the essence of paradigm shift is autonomous perception, autonomous decision-making, autonomous operation." Li Hongbo used Geek+'s own development experience as an analogy, saying the company's starting point was the goods-to-person model, using robots to bring shelves to pickers, helping pickers raise effective working time from 30% to over 95%. However, this is still weak AI and weak embodied intelligence with preset program automatic execution. The true leap occurs when robots can begin handling extreme operating conditions. Geek+ focuses on R&D and application of autonomous mobile robots, providing smart warehousing logistics solutions. The growth logic of the warehouse robot industry over the past decade was replacing repetitive labor with automation, with the technical premise that scenarios are sufficiently structured and tasks sufficiently predictable. The next decade's logic will become using agents to solve unstructured problems—this is not just an improvement in technical parameters.
李洪波还指出，当前具身智能行业存在重演示、轻落地的痛点。不少演示视频里，机器人流畅完成任务的画面，与真实场景中持续稳定运行之间还有着不小的鸿沟。“如果你不把端到端这个技术做透，就很难在7×24小时、一年365天持续给客户稳定交付。”从公司实操来讲，李洪波表示，传统AI机器人胜任的是单场景、少变化、简单关系的问题，而智能制造与智能仓储呈现的是跨场景、多变化、复杂关系的特点，必须突破具身智能、群体智能、数据智能三大关键技术，方能解决好能用、好用、用好的问题。
Li Hongbo also pointed out that the current embodied intelligence industry has a pain point of valuing demonstrations over deployment. In many demonstration videos, the smooth task completion images differ considerably from continuous stable operation in real scenarios. "If you don't thoroughly master end-to-end technology, it is difficult to continuously deliver stably to customers 24/7, 365 days a year." From the company's practical experience, Li Hongbo said traditional AI robots are competent for single-scenario, low-variation, simple-relationship problems, while intelligent manufacturing and smart warehousing present cross-scenario, multi-variation, complex-relationship characteristics. Three key technologies—embodied intelligence, swarm intelligence, and data intelligence—must be broken through to solve the "usable, easy-to-use, well-used" problem.
具身小脑解决能用问题。“一台仓储机器人以4米/秒的速度漂移过弯，在回南天地面打滑的条件下依然稳定运行。”李洪波表示，该机器人使用的漂移不是为了炫技，而是追求机械效率。在空间允许时以最高速度通过，减少无效停留。具身大脑解决好用问题。李洪波坦言，当前具身智能技术无法在更多样化的大场景里做大范围的跨场景学习。“比如，料箱被吹倒后，需要先扶正再捡起，姿势不对的货物需要先碰触调整才能搬运，这些不是预编程就能解决的。”他举例，这些都是具身大脑需要应对的典型极限工况。
The embodied cerebellum solves the "usable" problem. "A warehouse robot drifts through a corner at 4 m/s, still operating stably on slippery floors during humid return-south weather." Li Hongbo said the robot's drift is not for showing off but for mechanical efficiency—passing at maximum speed when space allows, reducing ineffective pauses. The embodied brain solves the "easy-to-use" problem. Li Hongbo admitted current embodied intelligence technology cannot do large-scale cross-scenario learning in more diverse large scenarios. "For example, after a material box is blown over, it needs to be righted before picking up; goods in wrong postures need to be touched and adjusted before handling—these cannot be solved by preprogramming." He cited these as typical extreme operating conditions the embodied brain must handle.
在解决了能用、好用的问题之后，群体智能与数据智能解决用好问题。活动现场，李洪波透露了极智嘉系统的极限规模，单个系统最大跑到3600台机器人。面对集中调度最优但不实时、分布式调度实时但非最优的问题，李洪波认为，解法是先基于模式学习，做全局最优调度，再通过分布协商做局部调度。而在数据智能层面，则可以通过实时优化货架热度分布，高频货架前移、低频货架后置等，大幅减少机器人搬运距离，提升整体效率。李洪波还提到，系统需要处理百万级SKU、千台级机器人以及大量IoT设备，数据智能的目标是让整个系统既能实时调度又能保持高效。目前，这套技术架构已在全球40多个国家商用近10万台机器人。
After solving the "usable" and "easy-to-use" problems, swarm intelligence and data intelligence solve the "well-used" problem. At the event, Li Hongbo revealed Geek+'s system extreme scale—a single system running up to 3,600 robots. Facing the problem that centralized scheduling is optimal but not real-time, and distributed scheduling is real-time but not optimal, Li Hongbo believes the solution is first doing global optimal scheduling based on pattern learning, then local scheduling through distributed negotiation. At the data intelligence level, real-time optimization of shelf heat distribution—moving high-frequency shelves forward and low-frequency shelves backward—can significantly reduce robot transport distance and improve overall efficiency. Li Hongbo also mentioned the system needs to handle millions of SKUs, thousands of robots, and large numbers of IoT devices. The goal of data intelligence is letting the entire system both schedule in real-time and remain efficient. Currently, this technical architecture has been commercially deployed with nearly 100,000 robots in over 40 countries worldwide.
随着具身智能的概念泡沫逐渐抹去，穿越周期要靠什么？“科技创新要回归原始创新，产业创新要回归商业的本质，通过双轮驱动给行业创造可感知的价值。”李洪波表示。
As the embodied intelligence concept bubble gradually fades, what does crossing the cycle rely on? "Technological innovation must return to original innovation, industrial innovation must return to the essence of business, and through dual-wheel drive create perceivable value for the industry." Li Hongbo said.
