Exclusive later丨Byte wants to build a robot team that plans to expand to 100 people

Original link: https://www.latepost.com/news/dj_detail?id=1731

“LatePost” exclusively learned that the byte robot team has about 50 people, and plans to expand to hundreds of people by the end of the year, and produce some robots that can serve byte’s own e-commerce fulfillment needs and can sort and pack goods in the warehouse.

The byte robot team is currently affiliated to the byte AI Lab, and its leader is Li Hang, the director of the AI ​​Lab. Li Hang served as the director and chief scientist of Huawei’s Noah’s Ark Laboratory. He joined Byte in 2017 and now reports to Yang Zhenyuan, vice president of Byte and head of algorithm technology.

ByteDance’s robot exploration began in 2020, when ByteDance founder Zhang Yiming showed interest in robots, and since then he will participate in robot project discussions from time to time.

About a month ago, Zhang Yiming, ByteDance CEO Liang Rubo, head of algorithm technology Yang Zhenyuan, ByteDance vice president of product and strategy Zhu Jun, and Zhao Pengyuan, who was once the head of ByteDance’s strategic investment and is currently the head of group strategy, participated in the and AI Lab robotics team discussion.

Different from the past, the purpose of this meeting is to discuss the industrialization direction and plan of Byte Robot. Previously, Byte’s investment in robots was mainly based on technical attempts. By the end of 2021, there will be only a dozen people in the team, and investment will gradually increase after 2022.

During this discussion, Zhang Yiming encouraged the team to set a bigger goal.

At the meeting, a member of the robotics team proposed that the commercialization threshold for the robotics industry is 1,000 units, which can be used as a reference for industrialization goals. The feedback from Zhang Yiming, Zhu Jun and some participating management staff was that “1,000 units is a bit small”: “Why does Tesla want to make 10 billion units, and we only make 1,000 units?”

Yang Zhenyuan mentioned in this discussion that there are no more than three possibilities for making robots: one is to pursue a pole flag on the top of the technology mountain, just like Boston Dynamics, to make a technologically advanced prototype; the other is to use existing relatively formed The solution serves the internal needs of Byte; the third is to use a relatively mature solution to find external customers and expand the scale.

Summarizing the thoughts of Byte’s management in this round of discussions, they believe that now that Byte is a robot, it should:

  • Integrate with existing businesses to serve good scenarios and industries;
  • Explore the combination of large models and robots, and pursue technological leadership;
  • The commercial value of humanoid robots needs to be further observed. A behavior like Boston Dynamics that has been unprofitable for 30 years may not be suitable for Byte.

Recently, the byte robot team discussed the business direction again, and the goal was clearer, which is divided into two parts:

One is to produce some robots, and give priority to the fulfillment needs of byte e-commerce; the other is to focus on cutting-edge technologies and explore the application of AI large model capabilities to robots.

E-commerce fulfillment refers to the process in which the platform delivers the goods to consumers after the e-commerce transaction is completed. It involves sorting, grouping and packaging in the warehouse, as well as logistics links.

Byte has now established some self-operated warehouses, which mainly serve Douyin Supermarket in Byte’s e-commerce business. The proportion of Douyin Supermarket in Byte’s e-commerce business is still very small. Byte wants to use robots instead of humans to complete the process of picking, handling and packaging. China’s labor costs are not expensive now, but there are other benefits to using robots: it can better cope with the surge in short-term warehousing workload brought about by big promotions and sales of explosive products.

From this, it is speculated that the robots that Byte will be working on may be mobile sorting robots that can deliver goods in e-commerce warehouses, and robotic arms with visual perception capabilities that can pack goods by themselves. In the industry, the former’s benchmarking products include Amazon’s Kiva, and domestic companies include Jizhijia, Kuaicang, etc.; the latter’s benchmarking companies include Mujin (Japan), Mecamand, XYZ, etc.

3fb5aafe2b95abae7c2293f3dd7de0f7.gif

The Jizhijia sorting robot delivers goods in the warehouse (left); the Mech-Mind 3D vision-assisted robotic arm picks up packages (right).

It may also be a new product and program. It is understood that byte has not yet framed the specific shape of the robot, and the shape ultimately depends on the requirements of byte e-commerce performance.

Byte has not yet set a specific target for the number of robots it will manufacture this year.

The increased attention and investment in robots may come from Zhang Yiming and Byte’s thinking on the new development of AI.

“LatePost” has reported that Zhang Yiming will often share with some Byte people recently his learning experience in AI papers and his thinking on ChatGPT. Starting from the large model, Byte has increased its investment in artificial intelligence research and development. This year, it has ordered more than US$1 billion of GPUs from Nvidia. GPU is an essential computing infrastructure for AI training.

Robots are now also considered to be an important direction for AI large models, and the two promote each other: large models can bring “common sense” to robots; robots that interact with the real physical world can provide new types of data for AI development.

Prior to this, Byte’s exploration of robots was relatively restrained: starting in 2021, the robot team under Byte AI Lab has tried to make building service robots, which can deliver meals and express packages in office buildings; I am also looking for some robots that may serve Byte’s e-commerce self-operated warehousing and logistics outside of Byte, but most of the solutions are not mature, and “ROI cannot be calculated”.

It can also be seen from the organizational structure of the byte robot team that it was not business-oriented in the past. Byte AI Lab, established in 2016, is currently divided into two groups: NLP (Natural Language Processing) and Research. The former provides technical support for Byte’s business, while the robot belongs to the Research group and does not directly support business.

Since 2019, Byte has also successively invested in some robot companies. For example, the sweeping robot company Yunjing Intelligent, Yinghe Robot founded by Shen Gang, the former deputy general manager of FANUC, Jiazhi Technology, Juxing Technology and Future Robot, which serve warehousing and logistics, cover household cleaning, distribution, logistics, industry, etc. and many other fields.

ea20c6532f4815661806b06266519a05.jpg

From left to right are the products of Cloud Whale, Juxing Technology and Future Robotics.

Most of these robotics companies were established around 2017. The commonality is to develop robots for specific scenarios and solve specific problems, so their products have different forms. This is because under the technical limitations of the previous period, robots cannot be truly universal. For example, the same robot can not only move parts in the factory, but also deliver food, and even help you wash fruits and chop vegetables.

After Tesla announced the Optimus Prime humanoid robot plan in 2021, the wind direction of the industry began to change.

The goal of Optimus is to replace humans with humanoid robots to perform dangerous, repetitive and boring tasks. In Tesla’s demonstration, Optimus has been able to water flowers, tidy up tables and carry items. At Tesla’s shareholder meeting this year, Musk said that the demand for humanoid robots in the future will be tens of billions, perhaps more than that.

9c46968b6edf56af9eb3f0d0aad50a5e.png

Tesla Optimus, unlike most current robots that use wheeled movement, Optimus uses a bipedal design that looks more like a human.

After Musk’s bold plan, the progress of large-scale model technology has added fire to the industry. Some are beginning to believe that universal robots may be coming sooner than expected.

At the ITF World 2023 Semiconductor Conference in May this year, Nvidia founder and CEO Huang Renxun mentioned in his speech that the next wave of artificial intelligence will be embodied intelligence (embodied AI), which can understand, reason, and interact with the physical world intelligent system.

Technology companies at home and abroad are also trying to combine AI models with larger parameters and robots to create more general robots.

OpenAI invested in a Norwegian humanoid robot company 1X Technologies in March this year, which is developing humanoid robots.

1484c27a7dca84c95c0a4dc8fbc63d58.jpg

1X Technologies has developed a wheeled dual-arm robot EVE (left), which uses a two-finger gripper for its hand (middle), and they are developing a bipedal robot NEO (right). From the conceptual diagram, the hand will use dexterous hands .

In June of this year, Google’s DeepMind released RoboCat, which applied the capabilities of large models to robots. DeepMind scientist Alex Lee said in an interview with TechCrunch: “We proved that a single large model can solve diverse problems on multiple robot entities. tasks and can quickly adapt to new tasks and robotic entities.”

84fcede72a57519330c685f4e4e57065.gif

Demonstration of DeepMind RoboCat, after learning human actions 1000 times, the robotic arm can complete object classification and picking.

Since the beginning of this year, Tencent RoboticsX robotics laboratory has also applied AI models to robot dogs; with the support of Alibaba’s Tongyi Qianwen large model, users can directly input commands in natural language in the DingTalk dialog box to remotely command the robot. In April, Xiaomi established Beijing Xiaomi Robot Technology Co., Ltd.; in June, Huawei established “Jimu Machine” to lay out robots and intelligent manufacturing.

How long the robot projects and companies started in the new round of upsurge can last remains to be verified. Building a robot team and steadily advancing research and development requires huge investment. The bigger challenge is that the data for training robots is much scarcer than the data for training general software AI systems.

OpenAI Chief Scientist Ilya Sutskever (Ilya Sutskever) was interviewed in April this year and was asked whether OpenAI’s abandonment of robots was the right decision? He said, “Yes, we really couldn’t continue to develop robots at that time.” If we want to collect the data needed to optimize robots, we need a large team to manufacture and maintain robots: “Manufacturing 100 robots is a huge investment, but Even then, you’re not going to get a lot of data.”

A robot practitioner told “LatePost” that the data collection of robots is similar to that of autonomous driving. The most direct way is to use real robots to do various tests to collect data, which is very costly; another important supplement is simulation technology. Build simulated environments and corner cases to capture data. There is currently not much high-quality public data available on motion control. Behind all this corresponds to the resources of technology and capital.

Byte’s past attempts at hardware were not as successful as its Internet applications. Dali desk lamp and VR headset PICO are precedents. Dali desk lamps no longer release new products, and at least half of the team of 1,000 people left at one time; PICO also experienced layoffs in the first quarter of this year, and lowered its sales forecast for 2023 to 500,000 units, which is 50% lower than the actual sales in 2022. Meta’s Quest2 will sell more than 7 million units in 2022.

Byte has no shortage of resources. It used to be good at doing miracles vigorously: after setting a goal, it mobilized all available forces such as manpower and funds to attack with all its strength in a short period of time. But the hardware industry and cutting-edge technology have another logic. It tests patience, resilience and cognition, and requires longer waiting and persistence.

This article is transferred from: https://www.latepost.com/news/dj_detail?id=1731
This site is only for collection, and the copyright belongs to the original author.