AI automatically generates robot control code, and a new blue ocean track has emerged

Fever.net reported (Text/Zhou Kaiyang) As one of the applications that most tests the applicability of scenarios and cost control, the market has opened up a new blue ocean track again with the further access to high performance. For example, humanoid robots have already shown a trend of accessing large language models, and we have seen similar demos at many industry exhibitions.

The hand-to-hand combat in the robot market is no longer just a competition between motion and visual perception systems. Whoever can grasp the dividend of AI computing power in the new era may be the first to grab the first piece of cake in this blue ocean market. The premise for granting this dividend is precisely the robot AI chip and the related software development platform.

Access to large language models, AI automatically generates code

As the leader in the field of AI chips, Jetson AGX in has been sought after by the robot industry ecosystem as soon as it was released last year. The computing power of 200TOPS@INT8 has increased several times compared to the previous generation of Jetson Nano and Jetson Xavier. The previous generation was still the mainstream product at that time and even now. Jetson's development ecosystem has more than 1 million and 6,000 companies. Xiaomi's Cyberone, Yushu's B1 and other products are using Jetson as a perception computing platform.

However, Jetson AGX Orin is still aimed at the industry-level market. In the face of consumer-level products, entry-level and mid-range products are still needed to replace the original Nano and Xavier. Therefore, NVIDIA immediately released the new Jetson Orin Nano and Jetson Orin NX. Depending on the configuration, the maximum computing power can reach 40TOPS and 100TOPS respectively.

The tool suites of Metropolis, Omniverse Isa and TAO, which were updated with Jetson AGX Orin, further reduce the time cost required for AI deployment, and can solve challenges in the robot development process such as perception and natural language understanding. Since NVIDIA began supplying in the fourth quarter of last year, a number of newly released robot products this year have used this chip, and its role is no longer just for perception computing.

Take the domestic robot Unitree as an example. The Unitree Go2 EDU version they released can be equipped with NVIDIA Jetson Orin high-computing power modules to provide an additional 40 to 100TOPS computing power. The lowest option is 8GB Jetson Orin Nano and 16GB Jetson Orin NX. Unitree Go2 is also connected to GPT, and it is more flexible when interacting with people, rather than a programmed response.

This may mean a cross-generational progress for companion robots. After connecting to GPT, all interactive activities will become humanized. After understanding the semantics of the instructions, the robot's control code can also be automatically generated through the code generation capability of GPT, thereby triggering more complex action instructions without additional requirements. At present, Jetson is the most comprehensive robot hardware platform that connects to large models, so it also provides support for domestic robot manufacturers to combine domestic large models.

AI customization requirements and ease of use of development platform

Although Horizon is a leading domestic chip manufacturer, this company with Robocs in its name is also steadily advancing the development of robots. Last year, Horizon launched the Horizon Hobot Platform, the first integrated hardware and software robot development platform in China, and the X3 Pie development board based on the X3 chip.

In addition to the X3 chip, the Hobot platform also includes the robot Together, robot reference Boxes, etc. X3 integrates the new Bernoulli 2.0 BPU architecture, which can provide 5TOPS computing power. While being compatible with general operating systems, TogetherROS optimizes the underlying X3 chip and improves the performance of some CV operators. In addition to reaching cooperation with software manufacturers, Horizon also works closely with many robot body manufacturers, such as Songling Robot and Zhikete Robot.

This year, Horizon launched a new RDK series robot developer kit. Based on this, the original X3 Pie was upgraded to RDK X3. In the first half of the year, a more integrated RDK X3 module was also released. The main frequency of the quad-core Cortex-A53 was upgraded to 1.Hz, which can meet the more customized needs of robot developer customers.

At the Horizon 2023 Robot Developer Creation Day in July, Horizon released RDK X3 2.0 and RDK Ultra. RDK Ultra is a high-performance, high-computing kit prepared by Horizon for robot development. Its end-side reasoning computing power can reach 96TOPS, and it is equipped with an 8-core Cortex-A55 CPU to provide high-performance processing capabilities. At the same time, in order to better reduce the adaptation and learning costs of developers, the hardware of RDK Ultra is even compatible with the Jetson Orin series development board.

At the same time, the RDK X5 kit based on the next-generation chip Rising Sun 5 is also under development and is expected to be released in 2024. Horizon also revealed some performance indicators of the platform, such as 8TOPS BPU computing power, and the 8-core Cortex A55 CPU will further increase the main frequency to 1.8GHz, and this time it is equipped with Mali.

At the Robot Developer Creation Day, TogetherROS was also upgraded. TogetherROS upgraded to version 2.0 introduces more algorithms and application examples, including a number of complex algorithms that have been realized due to computing power upgrades, such as BEV visual 3D perception, visual line patrol, and so on.

In order to develop robot applications more targeted, Horizon Robotics has also launched Node Hub, which nodes functions such as SLAM, target, positioning and autonomous navigation, so that different types of robots can complete application development by simply connecting these nodes in series, such as logistics robots that require SLAM, autonomous navigation and motion control, or sorting robots that require motion planning and target detection.

AI computing takes into account both reliability and real-time performance

. Another hardware platform that is often overlooked can also help the development of AI robots, that is the Versal AI Edge series of /. This series provides a platform that provides a higher AI energy consumption ratio than GPU through the combination of scalar engine, adaptive engine and intelligent engine. In general control embedded computing, the dual-core Cortex-A72 and dual-core Cortex-R5F are sufficient to meet the needs of applications and real-time control, while the adaptive engine is used to support the fusion algorithms of various sensors.

Its intelligent engine consists of an AI engine and an engine. Each AI engine consists of a 2D array of multiple engine blocks, and each engine block contains a VLIW single instruction multiple data vector with a main frequency of up to 1.3GHz, and is specially optimized for and advanced processing, which is sufficient to meet the low latency of real-time control of robots.

It is worth mentioning that the AI ​​engine provided by AMD is divided into two types, one is the AIE engine that balances ML and DSP loads at the same time, and the other is the AIE-ML engine that focuses on ML. The former has higher advanced signal processing capabilities, while the latter has higher machine learning inference performance. This is why Versal AI Edge adds an independent DSP engine to the intelligent engine, so that both deep learning in robot AI applications and signal processing in perceptual computing can be completed on a single platform.

However, from the perspective of use cases, the Versal AI Edge series is more suitable for industrial-level robots, such as collaborative robots, medical surgical robots, and so on. Such applications often require a larger number of heterogeneous sensor inputs than consumer-level robots, and have higher requirements for machine learning applications such as predictive maintenance and remote diagnosis. Moreover, judging from the configuration of its scalar engine and the support of its adaptive engine for TSN, it is also designed to achieve the fastest response speed under ROS in industrial, medical and other environments.

In conclusion

For the robot market, the pursuit of miniaturized, low-power, high-computing AI computing chips will inevitably be the general trend in the next few years. The matching chips and sensors have also started a new round of technological upgrades, with the goal of forming a perfect fusion of the perception layer, computing layer, and execution layer. From the perspective of technological accumulation alone, autonomous driving chip manufacturers are likely to be the first to reap the benefits, but many of them still need more energy to access different AI models in order to keep up with the new interaction trend.