2026可持续全球领导者大会于10月20日至22日在上海举行。在具身智能与未来产业主题论坛的圆桌对话环节，亮源新创联合创始人范廷翔围绕具身智能的落地门槛分享观点。范廷翔介绍，亮源新创主要做具身通用大脑，最终目标是让大模型这样的通用智能从虚拟世界走向真实的物理世界。在这一过程中，他发现目前存在两个门槛。
The 2026 Sustainable Global Leaders Conference was held in Shanghai from October 20 to 22. At the roundtable dialogue of the Embodied Intelligence and Future Industries themed forum, Liangyuan Xinsheng co-founder Fan Tingxiang shared views on embodied intelligence's deployment thresholds. Fan Tingxiang introduced that Liangyuan Xinsheng mainly does embodied general brains, with the ultimate goal of letting general intelligence like large models move from the virtual world to the real physical world. In this process, he found two thresholds currently exist.
其一是AI实体化，即硬件层面。范廷翔表示，让AI拥有可靠的硬件载体本身是一道门槛，而在资本推动下，上下游正让成本越来越低、可靠性越来越好，这道门槛正在被逐步跨过。他说，如果大家今天去观察人形机器人运动会，可以看到今年和去年相比，在小脑运控和硬件供应链上的能力有了大幅提升，今天机器人在硬件能力上已经有超越人的能力了。
The first is AI physicalization, the hardware level. Fan Tingxiang said letting AI have reliable hardware carriers is itself a threshold, and under capital's push, upstream and downstream are making costs lower and reliability better, and this threshold is being gradually crossed. He said if everyone observes the humanoid robot games today, they can see that compared with last year, capabilities in cerebellum motion control and hardware supply chain have greatly improved this year, and today's robots already have capabilities surpassing humans in hardware.
其二是算法层面。范廷翔指出，硬件门槛之外，算法缺少的仍是一种非常通用、非常灵活、能够应变不同陌生环境的能力。他透露，亮源新创在实验室内已可以让人形机器人做到在一个样板间完成非常多人类能做的家务，但要让模型进入万千家庭，对模型的泛化能力提出了较高要求。他说，这个门槛目前来说是卡在机器人能否进入到规模化场景的一个很大壁垒，这也是我们目前重点在研发的方向。
The second is the algorithm level. Fan Tingxiang pointed out that beyond the hardware threshold, what algorithms still lack is a very general, very flexible capability to adapt to different unfamiliar environments. He revealed that Liangyuan Xinsheng has already enabled humanoid robots in the laboratory to complete many household chores humans can do in a model room, but letting models enter millions of homes imposes higher requirements on model generalization. He said this threshold is currently a major barrier to whether robots can enter scaled scenarios, and it is also our current key R&D direction.
范廷翔介绍，亮源新创已开源一些通用的具身大脑，包括通用具身导航模型和通用具身全身控制模型，产业方使用这些模型不需要做微调，可以直接部署到场景里，通过一句话就可以让机器人做全身动作，或前往任何地方，能听懂人话也是非常重要的能力。但他坦言，模型如果永远在前沿实验室里开发，是很难达到真正的泛化的。他表示，希望与行业同仁共建生态，让模型进入更多场景。
Fan Tingxiang introduced that Liangyuan Xinsheng has open-sourced some general embodied brains, including general embodied navigation models and general embodied whole-body control models. Industry parties using these models do not need fine-tuning and can directly deploy them into scenarios, and one sentence can make the robot perform whole-body actions or go anywhere, and understanding human language is also a very important capability. But he candidly admitted that if models are always developed in frontier laboratories, it is very difficult to achieve true generalization. He expressed hope to build an ecosystem with industry peers and let models enter more scenarios.
这个判断放在前面讨论的产业脉络里，和几个信息能对上。前面讨论过北京人形数据与训练基地用真实场景做数据采集和模型训练，也讨论过国网天津电力把电缆隧道改成机器人学校，还讨论过上海推动8000台人形机器人进工厂做实训，以及英伟达和OpenAI同时押注物理AI、雪球讨论区在争哪个环节先赚钱。这些讨论反复出现的一个共同点是，硬件能力在快速进步，但泛化能力仍然是卡住规模化落地的关键。范廷翔说的两个门槛，硬件门槛正在被跨过，泛化门槛还在，和前面讨论的机器人能跳舞却夹不起鸡蛋、灵巧手不再拼自由度、工业机器人进厂做巡检和搬运但精细装配仍需迭代，是同一个问题的不同侧面。
Placed in the industrial thread discussed earlier, this judgment aligns with several information points. Earlier we discussed Beijing Humanoid Data and Training Base using real scenarios for data collection and model training, State Grid Tianjin Electric Power converting cable tunnels into a robot school, Shanghai pushing 8,000 humanoid robots into factories for practical training, NVIDIA and OpenAI simultaneously betting on physical AI, and Xueqiu discussion area debating which link makes money first. A common point repeatedly appearing in these discussions is that hardware capability is progressing rapidly, but generalization capability remains the key blocking scaled deployment. Fan Tingxiang's two thresholds, the hardware threshold being crossed and the generalization threshold remaining, align with the previously discussed robots that can dance but cannot pick up an egg, dexterous hands no longer competing on DOF, and industrial robots entering factories for inspection and handling while fine assembly still requires iteration, which are different sides of the same problem.
范廷翔提到亮源新创已开源通用具身导航模型和通用具身全身控制模型，产业方不需要微调就能直接部署。这个做法和前面讨论的OpenLoong开源体系、智元元苼生态发展计划、爱思科技Agnes Harness框架是同一个方向，都是通过开源和平台化，降低行业使用门槛。但他也承认，模型如果永远在前沿实验室里开发，很难达到真正的泛化。这句话点出了一个现实，开源模型能解决一部分复用问题，但泛化能力的提升，最终还是要靠真实场景的数据回流和迭代。这也是前面讨论的Real2Sim2Real闭环、Physical RSI、具身长记忆这些技术路线正在尝试解决的问题。范廷翔说希望与行业同仁共建生态，让模型进入更多场景，这个表态背后的逻辑是，泛化能力不是一家公司关起门来能解决的，需要足够多的真实场景和数据来喂养模型。
Fan Tingxiang mentioned Liangyuan Xinsheng has open-sourced general embodied navigation models and general embodied whole-body control models, deployable by industry parties without fine-tuning. This approach aligns with the previously discussed OpenLoong open-source system, AgiBot's Yuansheng Ecosystem Development Plan, and Agnes AI's Agnes Harness framework, all lowering industry usage thresholds through open source and platformization. But he also admitted that if models are always developed in frontier laboratories, it is very difficult to achieve true generalization. This sentence points to a reality that open-source models can solve part of the reuse problem, but generalization capability improvement ultimately relies on real-scenario data return and iteration. This is also what the previously discussed Real2Sim2Real closed loop, Physical RSI, and embodied long-term memory technical routes are trying to solve. Fan Tingxiang said he hopes to build an ecosystem with industry peers and let models enter more scenarios, and the logic behind this statement is that generalization capability cannot be solved by one company behind closed doors, and it requires enough real scenarios and data to feed the model.
