On July 18, 2026, the "RISC-V and Physical Intelligence Forum," co-hosted by the Shanghai Open Processor Industry Innovation Center (SOPIC) and VeriSilicon, was successfully held at the Sci-Tech Innovation Hall of the Zhangjiang Science Hall in Shanghai.
Dr. Wayne Dai, Chairman, CEO, and President of VeriSilicon, attended and delivered a keynote speech. He mentioned that the development threshold for the humanoid robot track is even higher than that of the automotive autonomous driving industry. This is because the constraints for autonomous driving are relatively clear: roads, obstacles, and traffic rules, with the core goal being "not to collide." In contrast, robot scenarios are generalized—picking up cups, helping the elderly, folding clothes—where the physical properties, spatial positions, and interaction methods of each object are different, leaving no unified rule boundaries to apply.
Figure | Dr. Wayne Dai, Chairman, CEO, and President of VeriSilicon
Humanoid Robots Are Currently at the L2 Level
If benchmarked against autonomous driving levels from L1 to L4, humanoid robots are currently roughly at the L2 level—capable of completing specific action sequences in controlled environments, but still showing a clear gap in the ability to reliably deliver tasks as seen in L3.
From a technical perspective, the current bottleneck in the robot track lies not in the cerebellum, but in the cerebrum. Although on the surface, multimodal large models and VLA (Vision-Language-Action) still show impressive scores on public datasets, their generalization ability drops sharply when deployed in real home environments. The fundamental reason is that LLMs (Large Language Models) learn language distributions, not the causal structure of the physical world.
"After encountering bottlenecks with Transformers, everyone is talking about 'world models' and 'physical AI,' but there is still no mature solution for differentiable simulators that can truly model physical laws," Dr. Dai emphasized.
In response, he shared two approaches: using video generation models for future frame prediction, which essentially remains the fitting of pixel distributions; or using reinforcement learning for policy search in simulation environments, but the sim-to-real gap remains significant.
However, he also candidly stated, "The so-called 'having interacted with countless people'—understanding human emotions, intentions, and social norms—is currently even further away and not within the scope of engineering feasibility."
The Architectural Logic of RISC-V
Against this backdrop, RISC-V, as a more flexible hardware architecture support, is highly anticipated.
In fact, the core advantage of RISC-V lies not in any specific technical metric, but in the layered design of its instruction set architecture. The more than 40 base instructions have been frozen and will no longer change; standard extension instructions continue to evolve and are shared by the community; custom instructions are defined by users, and each company can apply for patent protection.
Dr. Dai believes that the key to this layering is "leaving blank space"—reserving sufficient space for customization while ensuring the compatibility of the foundational software stack. In contrast, the problem with x86 and Arm is that their base instruction sets have bloated severely, carrying too much historical baggage, and the extension permissions are not in the hands of users. Some niche domestic architectures can achieve autonomy and controllability, but their ecosystems fail to take off. RISC-V, however, has the potential to achieve autonomy, controllability, and prosperity simultaneously, which is where its strategic value lies.
Consequently, the Shanghai Open Processor Industry Innovation Center (SOPIC) was established, positioning itself to build an open-source software ecosystem using an open hardware platform.
Open and Open-Source Are Two Different Things
Why emphasize "open" instead of "open-source"? Addressing this question, Dr. Dai explained, "The cost of open-sourcing hardware is extremely high—tape-out costs, EDA toolchains, and process library bindings cannot be solved simply by putting an RTL on GitHub. The RISC-V specification itself is an open standard, but the specific implementations are the core assets of each company."
The industry believes that the logic behind establishing such a patent pool is very practical. Because Arm can sue commercial RISC-V companies, but would not sue UC Berkeley, as the base specifications are public. However, for commercial companies implementing based on these specifications, the risk of infringement is real. Therefore, the role of the patent alliance is not to prevent competition, but to reduce the litigation risk across the entire ecosystem. Currently, the first batch includes about 300 patents in the pool, with 23 members.
Several Changes Worth Noting
Google has already implemented comprehensive support for RISC-V in AOSP (Android Open Source Project), not as an experimental branch, but in the mainline codebase. Following Arm's notification to terminate its architecture license agreement, Qualcomm acquired a high-performance RISC-V CPU company in December 2025—this is not a technical backup, but a hedge against supply chain risks.
At the industry level, the China RISC-V Industry Alliance was established in Shanghai in 2018, making Shanghai the first local government in China to explicitly support RISC-V with policies. On the education front, 15 universities and 8 enterprises jointly developed 600 pages of open-source courseware. From last autumn to this spring, courses have been offered in 13 universities, covering more than 1,600 students. The Dishui Lake Forum has promoted over 40 domestic chips across four consecutive sessions, with each session attracting over 15,000 online views. The 2025 RISC-V China Summit saw over 8,100 attendees from 17 countries, with 202 paper submissions and 120 accepted—the organizers did not have paid speeches, a detail that indirectly reflects the community's vibrancy.
Conclusion
The bottleneck for robots lies in the causal modeling of the physical world, which cannot be solved by simply scaling up LLMs. The value of RISC-V lies in the customization space left by its layered architectural design, allowing for differentiation without sacrificing compatibility. Currently, the industry's focus has shifted from "proving usability" to "ecosystem building"—university courses, patent pools, industry conferences, and software stack adaptation. These are all infrastructure-level tasks; they may not show explosive results in the short term, but they will determine competitiveness five years from now.