From 'Dumb' to 'Smart': The Intelligent Path of Service Robots

With the development of artificial intelligence and robotics technology, service robots are gradually transforming from machines that perform simple tasks to intelligent agents capable of understanding and adapting to complex environments. Among all categories of robots, they are the most hopeful and necessary to upgrade from “dumb robots” to “smart robots”.

In recent years, the concept of “smart robots” has become very popular. What kind of robots can be called “smart”? According to China’s GB/T 39405-2020 standard, “smart robots” are defined as robots that have stronger perception, learning ability, and autonomy, capable of adapting to more complex environments and task requirements.

From 'Dumb' to 'Smart': The Intelligent Path of Service Robots

According to our standards, robots are mainly divided into three categories:

Industrial robots, whose main functions include handling, loading and unloading, welding, spraying, processing, assembling, and cleaning;
Service robots, which are further divided into personal/family service robots and public service robots. The former mainly includes household chores, education, entertainment, elderly care, and assistance for the disabled, while the latter mainly includes dining, guiding, multimedia, and public entertainment;
Special robots, whose main functions include inspection and maintenance, search and rescue, patrol, reconnaissance, bomb disposal, installation, mining, transportation, surgery, and rehabilitation.

Among them, service robots have a particularly urgent demand for intelligent upgrades due to the complexity of their application scenarios, the closeness of their interaction with humans, and their enormous market potential.

China’s service robot market is still in its infancy, but the market growth rate is significant, and it is expected to continue rapid growth in the coming years. The application scenarios of service robots have expanded from finance, agriculture, and healthcare in 2018 to hotels/travel and restaurants in 2023, although they are still only deployed in some standardized scenarios.

In the past eight years (2017-2023), the service robot market has grown rapidly (source: IFR, TE Research Institute).

When exploring the future of service robots, we often ask: Are they simple “tools” or truly “smart” partners?

On May 17, at the 14th Songshan Lake China IC Innovation Summit Forum, a roundtable forum was held with the theme “Dumb Robots or Smart Robots?” Industry experts from different fields discussed the current situation and future of smart robots.

From left to right:

Host: Dai Weimin, Vice Chairman of the IC Design Branch of the China Semiconductor Industry Association, Founder, Chairman and CEO of Chipone Technology.

Roundtable Guests:

Mr. Chen Feng, Vice President of Rockchip Electronics Co., Ltd.

Mr. He Yunpeng, Founder and CEO of Chengdu Qiying Tailun Technology Co., Ltd.

Mr. Shi Qing, General Manager of Shenzhen Yijing Virtual Reality Technology Co., Ltd.

Mr. Wang Wei, Vice President of Product Marketing at Pengkan Technology (Shanghai) Co., Ltd.

Mr. Zhang Xiaodong, Chairman of the Wuzhen Think Tank.

Ms. Zhang Xiuyun, R&D Director of Xiaomi’s ecological chain.

Mr. Yuan Diwen, Chairman and CEO of Shending Technology (Nanjing) Co., Ltd.

Looking back to 2018, the main application scenarios for service robots were finance, agriculture, healthcare, warehousing and logistics, and companionship at home. Five years later, in 2023, the main application scenarios for service robots are still finance, agriculture, healthcare, warehousing and logistics, hotels/travel, and restaurants. Is there any technology that can enable service robots to achieve breakthroughs in application scenarios?

Dai Weimin pointed out at the opening that the emergence of ChatGPT in 2023 has brought intelligent robots from the single-function “dumb” into the era of strong intelligence. They differ from traditional robots in terms of perception, learning, and autonomy, and can adapt to more complex environments and task requirements. This roundtable focuses on service robots, particularly the rapid growth and potential of the Chinese market.

Question 1: Cloud Brain, Ideal or Fantasy?

Dai Weimin raised a core question: Is it feasible to place the “brain” of service robots entirely in the cloud using 5G technology?

Mr. Chen Feng believes that the brain of service robots cannot rely entirely on the cloud. Because service robots need to move in real-time from point A to point B or react to their surroundings in real-time, network interruptions would make it impossible to ensure service; local intelligence is indispensable. “Due to the risks of network disconnection, local intelligence and computing power are essential.”

Mr. Zhang Xiaodong also agrees with this view, but from the perspective of computer development history, he believes that “the trend is moving from centralized to distributed, and the democratization of cloud and edge computing in the future may bring new breakthroughs.” This hints at the potential of a cloud brain.

Experts generally believe that, on one hand, a cloud brain can provide powerful computing capabilities and storage space; on the other hand, network instability may become a bottleneck for its development. Although the cloud brain has its advantages, the combination of local intelligence and computing power will provide service robots with a more stable and flexible solution.

Question 2: In the Next Three Years, Which Service Robots Will Enter Thousands of Households?

Cleaning robots are currently the most widely used and numerous service robots. Apart from that, in which fields are service robots most likely to achieve widespread adoption?

Dai Weimin envisions the market prospects for service robots, particularly highlighting the importance of companion robots for children’s education and elderly care. He believes these robots will become indispensable members of families, providing emotional and daily support.

When discussing consumer service robots that can be mass-produced in the next three years, Dai Weimin emphasized the potential of AI companions, especially in children’s education. “The early years are crucial for developing learning habits,” he said. AI companion robots can not only become good partners for children but also help parents better understand and nurture their children through data analysis.

At the same time, he mentioned the market demand for cooking, elderly care, rehabilitation robots, and entertainment robots, pointing out that comprehensive function robots have broad potential.

Question 3: Driven by Large Models, the Catalyst for Smart Robots?

The evolution of smart robots can be divided into several stages, from L0-level non-intelligent that relies entirely on human manipulation to L4-level super intelligence that completely replaces humans. Currently, most service robots are between L1 and L3 levels, among which:

L1: Basic intelligence, with certain autonomous learning capabilities, but limited decision-making ability.

L2: Moderate intelligence, with higher autonomous learning capabilities, but still requires human intervention at critical moments.

L3: High intelligence, with strong autonomous learning and decision-making capabilities, but cannot continuously self-learn or self-optimize; in some cases, still requires human assistance.

With technological advancements, smart robots will develop towards higher levels of intelligence in the future, reaching L4 level—super intelligence, with extremely high autonomous learning and decision-making capabilities, able to perform tasks in extremely complex environments and completely replace humans.

The introduction of large models has brought revolutionary changes to the robotics industry. From RNN, LSTM to Transformer, and then to the successful application of large models like BERT and GPT, advancements in natural language processing technology have brought a qualitative leap in the understanding and interaction capabilities of service robots. Experts predict that large model-driven robots will hit the market in the near future, bringing more convenience to human life.

Compared to differentiated multimodal large models, language-based multimodal large models will become mainstream. The progress of natural language processing has undergone five paradigm shifts:

1950-1990: Small-scale expert knowledge.

1990-2010: Shallow machine learning algorithms.

2010-2017: Deep learning algorithms.

2018-2023: Pre-trained language models.

2023-present: ChatGPT.

The development context of large models can be summarized in the following stages:

Early stage: Recurrent Neural Networks (RNN) and its variant Long Short-Term Memory (LSTM) were the main methods for processing sequential data.

Key year of transformation:2017 marked a turning point when the Google Brain team introduced the Transformer model, which incorporated the attention mechanism (self-attention), effectively addressing the limitations of RNN in sequence processing, and quickly became the mainstream model in the field of natural language processing (NLP).

The Rise of Transformer:Subsequently, the emergence of large models such as BERT and GPT further validated the flexibility and strong performance of the Transformer architecture, achieving significant results in various NLP tasks such as machine translation and time series prediction, and promoting the popularization of pre-training and fine-tuning paradigms.

Scale Competition:Entering 2021, the scale of models continued to expand, with Google’s Switch Transformer marking the industry’s entry into the trillion-parameter era, demonstrating the demand for large-scale computing resources and more efficient architectures.

A New Chapter in Interactive AI:In November 2022, OpenAI launched ChatGPT, a natural language processing tool based on advanced AI technology, revolutionizing human-computer interaction experiences and showcasing the immense potential of AI in dialogue understanding and generation.

These series of advancements reflect not only the iteration of technology but also mirror the active role and fierce competition of tech giants like Google Brain, OpenAI, Google, NVIDIA, and Microsoft in advancing the frontiers of artificial intelligence.

At the 13th Chipone CEO Forum, experts predicted that by 2028, sales of edge fine-tuning cards and inference cards will exceed those of cloud-side training cards. This means that after the “roots” (computing power) and “trunk” (general large models, training cards) are fully developed, the “branches” (domain-specific large models, fine-tuning cards) and “leaves” (applications, inference cards) will flourish in the coming years.

At this roundtable forum, guests and on-site audience members generally agreed that large model-driven robots will be on the market within three years. In fact, similar products have recently emerged, such as the Unitree G1 humanoid intelligent robot released by Hangzhou Yushu Technology, equipped with the UnifoLM (Unitree Unified Large Model), a powerful AI technology platform that enables robots to self-learn and upgrade iteratively.UnifoLM, as one of the core driving forces of G1, not only enhances the intelligence level of robots but also provides a co-creation platform for the application development of humanoid robots, promoting the development and innovation of intelligent agent technology.

Mr. He Yunpeng believes that large models are very important for service robots, especially for general service robots that need to understand world knowledge. However, for specific tasks, such large model parameters may not be necessary. We should solve specific problems based on application needs, combining traditional AI algorithms.

Mr. Zhang Xiaodong succinctly expressed his views on large models: “With a large model, there is intelligence; without a large model, there is dumbness.” He believes that the emergence of large models has greatly promoted intelligence development, and the trend of miniaturizing large models and sinking them to the terminal will make large models more popular. Today’s AI smartphones and AI PCs are the best examples.

Question 4: Humanoid Robots, Future or Past?

Humanoid robots, as typical representatives of embodied intelligence, have attracted widespread attention in the industry. NVIDIA CEO Jensen Huang has emphasized the importance of embodied intelligence on multiple occasions and predicted that humanoid robots will become mainstream products in the future. For example, at ITF World 2023, he pointed out that embodied intelligence is the next wave of artificial intelligence, where such intelligent systems can understand, reason, and interact with the physical world.

At the GTC AI Conference, NVIDIA released models and chips for humanoid robots, calling it “NVIDIA’s soul.” Jensen Huang predicted that the manufacturing costs of humanoid robots would be far lower than expected, making them a popular product.

The best example is the Unitree G1 humanoid intelligent robot mentioned above, released in May this year, with a starting price of 99,000 yuan. Data shows that Unitree Technology focuses on the independent research, production, and sales of high-performance general-purpose bipedal/humanoid robots and dexterous robotic arms. They have been invited to participate in major events such as the 2021 CCTV Spring Festival Gala, the 2022 Winter Olympics opening ceremony, the 2023 Super Bowl pre-performance, and the 2023 Hangzhou Asian Games and Asian Para Games. They are also the world’s first company to publicly retail high-performance quadrupedal robots and have achieved early industry implementation, leading global sales year after year.

Various parties are optimistic about the future development of humanoid robots, making them gradually become the new favorite in the service robot market. So the question arises: Do service robots necessarily have to be humanoid?

“All researchers in humanoid robotics share a common view—humans are the most efficient organisms. However, my view is somewhat extreme; I believe that human (humanoid) bodies are not as capable as other organisms.” A simple example is that when climbing a wall, humanoid bodies (with two hands and two feet) are not as effective as spider-like bodies (with eight legs). Zhang Xiuyun believes that service robots do not have to be humanoid, but humanoid robots can provide more natural and friendly interaction experiences, possibly because of empathy and emotional reasons, making them more easily accepted by humans.

Yuan Diwen added that the development of humanoid robots is closely related to the progress of large models, as they can better simulate human activities and perspectives, playing a key role in data collection and humanoid robot development. He gave an example: before 2020, the capital market was very resistant to bionic (humanoid) robots, but after the emergence of large models, the situation changed significantly.

“Now, the development of humanoid robots is actually driven by large models. Many large model companies are actively collaborating with humanoid robots for data collection.” Yuan Diwen believes that since humans can do many things, how robots achieve human-level capabilities should be trained from a human perspective. “The robotics industry needs such large models.”

The technical framework of intelligent robots (source: “White Paper on the Development of Intelligent Robot Technology Industry (2023 Edition)”).

Mr. Shi Qing also believes that service robots do not necessarily have to be humanoid, but humanoid design helps robots better integrate into human society and family environments. For example, a recent video released by Musk showing a robot was trained by a trainer wearing a VR headset to perform actions, where the robot learns. “The reason for training in a digital twin manner is to make the robot more human-like; conversely, if training a spider-like robot, then the application scenario may not be suitable for families, and the intimacy level would be compromised.”

Question 5: Optical Fiber Communication, Making Robot Reactions More Timely?

With the rapid development of large models like GPT, Transformer, and Wenxin Yiyan, bringing the typical representative of embodied intelligence, “humanoid robots”, back to the center stage is only a matter of time. However, we must also recognize the pain points brought by increased demand:

Super complex motor control: Robots have 30-50 or more joints.
Super numerous sensor access: Robots need to connect to an increasing number of sensors.
Delicate appearance design: Robots are developing towards more miniaturization, with limited space.
Bandwidth delay: Upgrading Ethernet solutions to over 1G networks is a significant challenge, with high latency and packet loss rates.
Communication standards: Multiple protocols coexist, causing a lack of uniformity.

Experts believe that future service robots will require higher-performance MCUs, more powerful NPUs, and more efficient communication networks.

Figure 1: Common independent networks for robots. Figure 2: TS-PON network integrating communication and control.

Mr. He Yunpeng pointed out that the chips required for robots are not just one or two, but a series of chips used to complete different tasks. It may even require different types of computing chips to collaborate on combined tasks. If we compare robots to human characteristics, they need a brain, limbs, cerebellum, and sensory nerve endings. The robot’s chips interact with humans based on task division. After obtaining the target task, it needs to understand the human’s purpose and intention, then decompose the task into subtasks and further think about how to implement and plan. “Task decomposition belongs to the brain’s tasks; after decomposition, it is given to the cerebellum to be responsible for action, which controls the robot’s posture, emergency avoidance, etc.”

“Cerebellum chip” or “neuron chip” is equivalent to an edge computing chip, where low-power design is crucial for this type of chip. Integrating AI into sensors, “perception and computation in one” can significantly reduce power consumption. He Yunpeng said, “However, some connection chips are still needed to link the data together because perception is distributed.”

Mr. Wang Wei proposed an innovative idea of using optical fiber (TS-PON solution) as the communication network for robots. He believes that using a single optical fiber to carry all electrical bus services provides a high-bandwidth, low-latency, and interference-resistant communication method, which is expected to become the upgrade direction of communication networks for service robots.

Mr. Wang introduced the case of Pengkan Technology using optical fiber communication networks to replace traditional cables, successfully solving issues of bandwidth, reducing weight, and electromagnetic interference. It is reported that this technology has been applied in advanced humanoid robots.

TS-PON network in humanoid robot systems.

The advantages of Pengkan’s TS-PON (Time-Sensitive Passive Optical Network) solution include:

Bandwidth/Delay:Currently, single-fiber 10G, soon to be 50G; delay at the µs level, no data loss, and no EMI/EMC interference.

Super high-performance MCU:Fully leverage the performance of high-performance AI processors and advanced sensors, effectively enhancing flexible control capabilities.

A single optical network:A single optical fiber carries all electrical bus services and supports up to 128 nodes, making wiring very simple.

Multi-protocol fusion network:Flexibly adapts to different hardware interfaces and protocol stacks, unifying standards, integrating resources, and breaking through fragmentation bottlenecks.

In the face of emerging industrial demands, such as autonomous driving, flexible production, and the rapid development of high-precision robotics technology, higher requirements are placed on industrial networks, especially in terms of high bandwidth, low latency, and strong anti-interference capabilities to support the increasing number of sensors and enhanced AI processing capabilities. Traditional copper-based industrial network technologies have gradually struggled to meet these new standards, constituting a “bottleneck” problem for industry development.

In this context, Pengkan’s TS-PON integrated sensing and control chip utilizes the advantages of optical communication to replace traditional solutions, meeting the urgent needs of automotive manufacturers, robot manufacturers, etc., for high-speed, low-latency connections. With domestic chip manufacturing processes reaching 14nm and even more advanced, developing higher-performance SoCs is just around the corner.

Pengkan TS-PON integrated sensing and control SoC chip architecture diagram.

These SoC chips integrate multiple functions, not only improving the performance indicators of the industrial internet, including accuracy, processing power, and energy efficiency, but also reducing costs and complexity through highly integrated design, thus providing key technical support for the comprehensive upgrade of industrial networks and completely solving the “bottleneck” problem, promoting industrial upgrading.

Yidao has also been focusing on the AR/VR track in recent years, which has a high degree of compatibility with the robotics industry. Mr. Shi Qing believes that optical fiber communication has great potential in the communication networks of robots. It can solve communication issues between robot joints and has advantages in anti-interference and high bandwidth.

Conclusion

The future of service robots is full of infinite possibilities. From cloud brains to large model-driven technologies, from humanoid designs to innovations in terminal chips and optical fiber communication, each step of development is a rethinking of the definition of “intelligence”. With the continuous advancement of large models and AI technology, terminal devices need higher-performance chips to handle complex tasks. This includes natural interaction, task decomposition, behavior planning, and perception processing.

Through this roundtable forum, we can see that the service robot industry is standing at a new starting point. Dai Weimin stated that we have reason to believe that service robots will no longer be simple tools but intelligent partners that can truly understand and serve human needs.

In this leap from “dumb” to “smart”, we look forward to service robots bringing more surprises and possibilities.

Recommended Hot Articles

Men Remain Boys Until Death: The First SMT Solder Paste Machine in the Hands of Older Men
Disassembling a 9.9 Yuan Shipping Charging Button Night Light
Sharing Several High-Quality Free Circuit Design Software
Disassembling a Handheld Thermometer: Cost-Cutting Operations Everywhere
Jumping Jobs in the U.S. Is Innocent; When Will Domestic Follow Suit?

Related posts

Leave a Comment Cancel reply