Code View | Chen Runze, Source Code Capital: New Changes, New Signals, New Directions — How Is AI Transforming Robotics?

Recently, Chinese tech think tank Jazzyear hosted its 2023 Jazzyear Gravity Year-End Gala under the theme "To You, Chasing the Wind and the Moon." At a roundtable titled "How AI Is Transforming Robotics," Runze Chen, Executive Director at Source Code Capital, shared his thoughts on new shifts in the robotics industry, early signs of AI-driven change, and evolving investment strategies.

He believes that new AI technologies, particularly large models, are altering the fundamental characteristics of robots. Robots have made significant strides in perception and understanding — a major transformation for both robotics and autonomous driving. Human-robot interaction has also advanced considerably, the most direct application of large language models. As for physical-world interaction, data and model scaling methodologies show strong potential, though the maturity of these technologies needs careful evaluation.

On the investment front, he suggested two possible approaches: one, capitalizing on cyclical opportunities in downstream industry investment; and two, tracking the continued progress of general-purpose robots in capability and cost.

01

New Shift: General-Purpose Intelligent Robots Become This Year's Hot Topic

Robotics investment cooled somewhat between 2021 and the first half of 2022, likely due to a phase shift in the supply-demand dynamics between capital and startup assets. But by late 2022, several developments began pushing the industry forward: first, Tesla's humanoid robot; second, teams at Google and elsewhere began demoing impressive large model + robotics integrations. This drove a "narrative upgrade" in the sector. Entering 2023, as general-purpose intelligent robots gained attention, more people began recognizing the potential of reinforcement learning, large models, and diffusion models in robotics — and market sentiment kept heating up. But we should be mindful that some of this may contain irrational elements.

Looking back at robotics projects from recent years, the main technical variables have been localization/navigation and computer vision. Many mobile robot categories are already fairly mature, with AI applications concentrated on object recognition, detection, and localization/navigation. Most startups have essentially been using these technologies to find product-market fit in specific scenarios. But we might broaden our perspective: looking at the past decade of automation industry growth, we shouldn't focus solely on supply-side changes — we need to watch downstream developments too. Some of the better, larger companies that have listed on A-shares all have strong industry-specific attributes. There's demand within an industry, and robot technology happens to address that demand. These are often the best opportunities.

02

Early Signs: AI Begins Transforming Robotics

Over the past year or so, advances in AI have given us greater confidence in robotics. However, customers don't actually care whether you use AI or not. In some manufacturing scenarios, clients only care about two things: first, whether you can meet performance requirements for cycle time, precision, throughput, and so on; second, whether the cost makes sense — and cost includes not just the robot itself, but also a substantial portion of deployment costs. We've spoken with many robotics entrepreneurs, researchers, and engineers, and we clearly sense that recent progress in vision-language models has brought qualitative improvements to robots' perceptual understanding. This advancement will drive rapid change in both robotics and autonomous driving. When we place robots in open-world environments, they can now understand the semantics of the world — fundamentally different from how robots were built before. Previously, machines understood the world through recognition of specific objects. Take autonomous driving: the perception module's whitelist used to be very short, but today's semantic segmentation and understanding capabilities have expanded dramatically, which has major implications for downstream planning. Or consider inspection scenarios: previously requiring extremely detailed annotation, but today deployment costs for such scenarios could drop significantly.

On the question of robot interaction with the physical world, we're closer than ever to finding a scalable, generalizable approach to robot learning. For certain domain-specific tasks, we believe the technology has reached the engineering stage. But general manipulation capability remains, at least in the near term, a luxury we can't yet expect. Admittedly, there's been excellent academic research on robotic manipulation recently — we've seen lab robots capably handling rigid objects, and even manipulating deformable objects like clothing and plastic bags. But bringing these technologies out of the lab takes time, and requires industrial forces to push the flywheel forward.

Compared to physical-world interaction, human-robot interaction benefits more directly from large language model development, and we'll soon see commercial deployment opportunities. But it's worth emphasizing: once physical-world interaction is involved, we need to carefully judge whether the technology is mature enough and whether commercialization can proceed smoothly.

Autonomous decision-making is also a critical component of general-purpose robots. We believe large language models have laid good groundwork for this, but the technology's maturity remains to be seen.

We strongly feel that AI + robotics has reached a crossroads where academia and industry must join hands. Source Code Capital hopes to collaborate with entrepreneurs to identify launchable scenarios, get data flywheels spinning, and explore appropriate hardware forms through real business operations — and on that foundation, advance the development of robotics + large models.

03

New Investment Direction: Focus on Whether R&D Can Translate to Real-World Deployment

When evaluating investment opportunities in robotics companies, two main threads deserve attention. The first is tracking capital investment trends in downstream industries and whether a company can seize opportunities and align them with its technology. Some excellent robotics and automation companies — Inovance, Supcon, Maxwell, Lead Intelligent Equipment, and NAURA, among others — all show distinct patterns tied to downstream industry capital cycles. Before truly general-purpose robots arrive, we believe robotics must still be viewed against a backdrop of large-scale capital investment.

The second thread is generality. Source Code Capital didn't rush to invest after humanoid robots became trendy, but we do recognize the need to find a maximum-common-denominator hardware form that can drive down costs through mass production and manufacturing. On that basis, robot application development should become as close to software R&D as possible — a path validated by PCs, smartphones, and other devices. So when assessing general-purpose robot investments, we need to see whether companies are doing substantial work around reliability, manufacturing, and cost. Of course, any hardware form factor convergence also depends on scale; even general-purpose robots need to find a suitable industry with sufficient capital investment. So both threads ultimately come down to demand.