CEO Tips丨When Satellites Meet AI, Does the Satellite Have a "Brain" from Then On?

Recently, there are several hotspots in the aerospace field. Not only did SpaceX's Starship 5 "rocket held by chopsticks" succeed, but China also successfully launched the world's first AI satellite. After AI enters space, how to make satellites more "intelligent"? What conveniences might we enjoy from the AI satellite technology in the future? And is Elon Musk's Mars plan going to be advanced?

At 7:00 PM on October 16, the 36Kr Live Studio invited Tian Feng, the dean of the SenseTime Intelligent Industry Research Institute and the founder of "Tian Feng Says", Wei Wenyi, the CEO of Silk Road AstroMap, and Niu Min, the founder and CEO of Future Aerospace, to join us and discuss what sparks will be generated when satellites meet AI.

In this live broadcast, the three guests mainly discussed the following issues:

1. On September 24, 2024, the world's first AI large-model scientific satellite was successfully launched into orbit. What is the significance of the successful experiment of this AI satellite for satellite technology?

2. Satellite applications already need to consider energy consumption, and large models are even more computationally intensive applications. How to balance the energy consumption problem between AI computing power and satellite applications?

3. The value of AI is indeed great. How does AI make satellites more intelligent in space?

4. The data collected by satellites in space is huge and complex. After the addition of AI, will there be any differences in the data acquisition of satellites?

5. Currently, the application of commercial aerospace in the C-end market is relatively blank. Does the successful verification of AI satellites open up new business opportunities for us?

6. Satellite applications are in a slow growth stage. In the current market environment, should commercial aerospace companies focus more on deepening the proven commercial applications or actively explore unknown new applications to find more growth points?

36Kr: On September 24, 2024, the world's first AI large-model scientific satellite was successfully launched into orbit. What is the significance of the successful experiment of this AI satellite for satellite technology?

Tian Feng: The long history of satellite technology and the early ideas of Mr. Qian Xuesen on satellites and artificial intelligence. Since the Soviet Union successfully launched the first satellite in the 1950s, the satellite technology competition between the United States and the Soviet Union has begun. Mr. Qian Xuesen had foreseen the concept of a "stellar dock" (i.e., a space station) composed of satellites and a "fifth-generation computer" or "intelligent machine" integrated with the aerospace system in that era. With the development of the times, AI technology has gradually matured and achieved a deep integration with satellite technology. This integration has significantly improved the autonomy and task coordination ability of satellites. The operating environment of satellites in space is complex and changeable, and the signal transmission delay is long, making real-time ground operations difficult. Therefore, satellites need to have a certain degree of autonomy to deal with situations such as communication delays or failures. The application of AI large models provides this autonomy to satellites. When the connection between the satellite and the ground is slow or it needs to handle complex new tasks, the AI large model can guide the satellite to make autonomous decisions and adjustments to ensure the successful completion of the task. In addition, through AI technology, multiple satellites can form a constellation alliance and use swarm intelligence for complex operations. This collaboration not only improves the working efficiency of satellites but also enhances the overall stability and reliability of the satellite system.

Wei Wenyi: In the current aerospace field, the application of AI technology is changing the way data is processed. The traditional satellite data processing process requires a lot of manual participation, including data reception, preprocessing, and annotation. These tasks are not only time-consuming and labor-intensive but also inefficient. However, with the introduction of AI technology, this situation has been significantly improved. AI technology automates and intellectualizes the data processing, making the originally manual data processing processes streamlined and mechanized, thereby greatly improving the data processing efficiency. Wei Wenyi mentioned that now, from data reception to preprocessing, and to the final analysis and interpretation, the entire process can form a streamlined sorting model, reducing the data processing time from several days or even weeks in the past to several hours. In addition, AI technology also solves the problem of multi-modal data fusion processing. In the process of satellite earth sensing, it is necessary to fuse multi-source data from the ground, air, and space. However, these data often have different formats, resolutions, and accuracies, and it is difficult for traditional processing methods to achieve effective fusion. However, AI technology can achieve precise fusion of multi-modal data by learning and understanding the internal laws and correlations of the data, thereby providing more comprehensive and accurate information.

Niu Min: Traditional manual annotation work is not only labor-intensive but also requires professional knowledge. This is especially true for remote sensing data annotation, as it requires the identification and understanding of various geographical features and objects. However, with the development of AI technology, this cumbersome and complex work is being gradually replaced. AI can automatically identify and annotate the features in remote sensing data through learning and training, thereby greatly improving the data processing efficiency and accuracy. Satellites are actually more like a "vehicle" with a power system. It has its own "brain" (i.e., control system), but currently, this "brain" is more like a "cerebellum", mainly responsible for controlling the satellite's attitude and orbit adjustment. However, with the introduction of AI technology, the "brain" of the satellite will become more intelligent and powerful. The main application of AI on satellites is in the processing of remote sensing information, which is similar to edge computing in intelligent monitoring. By deploying AI on satellites, real-time data processing and decision-making can be achieved, thereby reducing the delay in data transmission and processing. In the future, as more satellites are launched and form a constellation network, inter-satellite connections and the formation of clouds will become crucial. The computing power constellation composed of AI satellites will have greater practical application value.

Tian Feng: Let me add a small story. The first scenario of SenseTime in the field of AI remote sensing is that the client company bought a large amount of satellite photo data, but many of the photos could not be used due to cloud coverage. Therefore, the first algorithm developed at that time was used to identify and filter out these cloudy photos so that the remaining photos could be used for further analysis. Some photos can be processed to remove the clouds through algorithms, but some cannot be processed due to the large area of cloud coverage.

Mr. Niu also mentioned two concepts of remote sensing technology in observing Earth objects. The first is the perception object. Previously, satellites might only be able to identify specific objects, such as ships. But now, large AI models have the ability of ubiquitous intelligence and can identify various objects from tower cranes to buildings, to hydrology, forests, and even large animals. This large AI model can be regarded as a general perception ability, which is very valuable. Because these models can be trained on the ground and then the number of parameters can be reduced, for example, compressing a large model with 1 billion or several hundred million parameters into a small model suitable for satellite carrying.

The other important concept is identification and classification. For example, identifying a scenic area or a building requires classification and marking with different colors. In the past, this kind of classification and annotation work needed to be done manually, and the workload was very large. But now, with the arrival of the era of large models, large models can automatically perform annotations. For example, plants can be marked with green, hydrology with blue, and road networks with purple, so that the layers and objects can be quickly classified. This greatly facilitates the subsequent processing work. In general, the application of AI technology in the field of remote sensing not only improves the efficiency of data processing but also expands the perception ability of satellites, enabling satellites to identify and classify more Earth objects.

36Kr: Satellite applications already need to consider energy consumption, and large models are even more computationally intensive applications. How to balance the energy consumption problem between AI computing power and satellite applications?

Niu Min: The development of satellite technology is similar to the evolution of the automotive industry. Just like the transition from fuel vehicles to electric vehicles, satellite technology is also undergoing an energy integration revolution. In the past, many satellites relied on chemical propellants to adjust orbits and attitudes. The lifespan of satellites is often not due to hardware damage but the exhaustion of fuel, resulting in the inability to maintain the orbit and failure. This is similar to the limitations of fuel vehicles, which require regular refueling.

Now, satellite technology is developing towards an electric propulsion system, which is similar to the electric motor of an electric vehicle. The electric propulsion system requires a power supply to maintain the orbit, which requires the satellite to be equipped with an efficient energy system. For computing power satellites, energy consumption is an urgent problem to be solved. In space, solar energy is the only energy source, but when satellites in low Earth orbit orbit the Earth, there will be a period of time in the Earth's shadow, and at this time, the battery energy carried by the satellite needs to be consumed. Currently, many satellites carry lithium batteries to provide energy storage.

In the future, two technical solutions may be adopted. One is to further reduce the power consumption of satellites, and the other is to carry out energy technology innovation, such as using perovskite solar cells or developing energy storage batteries with higher energy density. In addition, with the emergence of computing power satellites, in the future, new types of spacecraft specifically for energy storage may appear, similar to power satellites. These satellites can be responsible for collecting solar energy and storing it, and then providing energy to high-energy-consuming satellites through laser or microwave transmission.

In general, the future development of satellite technology will be in the direction of reducing energy consumption and improving energy efficiency, which has a similar trajectory to the transition from fuel vehicles to electric vehicles in the automotive industry.

Tian Feng: The computing power problem mentioned by Mr. Niu is very interesting, and it is indeed a knowledge blind spot for many of us. Now the computing power of satellites is already very powerful. For example, if the computing power of a robot is 100T, then the computing power of an autonomous vehicle can also reach 100T. The computing power of satellites can even reach more than 200T, and the latest satellites can be configured with 1000T of computing power. Some satellites are even named after computing power, showing their strong computing capabilities.

However, whether such a large computing power can be powered by the existing lithium batteries is a problem. Here we can think from another perspective, that is, the development of chips conforms to Moore's Law. Moore's Law states that approximately every 18 months, the performance of chips will double, while the cost will decrease. This means that compared with this year's 200T computing power chip, the cost and energy consumption of next year's 200T computing power chip may be lower.

Therefore, we can expect that perhaps within the next one or two years, large AI models will become more energy-efficient, while the computing power of chips will be stronger, but their energy consumption and cost will not increase significantly. This is actually the exponential dividend brought to us by Moore's Law. In general, with the progress of technology, we can expect that the computing power of satellites will continue to increase, while energy consumption and cost will be better controlled. This will bring great potential and opportunities for the development of satellite technology.

Wei Wenyi: What we just mentioned is how to use computing power on the satellite. In fact, not only on the satellite, the computing power required for the application of AI in the field of remote sensing is very large on the ground. For example, I just encountered a case today. There is an agricultural large model that needs to calculate the entire production cycle of a certain crop in all counties and cities across the country. This is itself a huge challenge. In the field of remote sensing, each crop has its unique regional characteristics, so it is not possible to simply apply a model from one region to another. This means that if a large model is to be used to uniformly handle these problems, a very powerful coupling mechanism is required, which is already very difficult on the ground.

Take the Xidian-1 satellite as an example. This satellite was launched in 2022 and is equipped with a hyperspectral and model-defined payload. These payloads are used for target recognition and contour extraction. For visible light images, they can identify targets; for hyperspectral satellites, they can sense and identify the spectral characteristics of different crops. Because the amount of data in each scene is very large, about 10GB, the pressure in the data transmission process is very high. In order to reduce this pressure, the satellite is equipped with an on-orbit compression payload, which can directly compress the data on the satellite and then transmit the compressed data back to the ground, thus greatly reducing the transmission pressure. If the AI large model can be directly deployed on the satellite, then the satellite can directly process the data and transmit the processing results, which will be more convenient. Or, only the eigenvalue is transmitted, so the amount of data will be much smaller. In this way, the satellite can help us filter out the required data without processing useless data.

In addition, the sensors on the satellite are also very important. For example, infrared sensors can directly see things underground and are of high value. Devices like SAR satellites can penetrate clouds and rain and are very useful for remote sensing applications. In general, with the development of technology, we can expect that the sensors and AI models on the satellite will be more efficient and intelligent, and can better meet our needs.

36Kr: The value of AI is indeed great. How does AI make satellites more intelligent in space?

Niu Min: Let me give an example of an autonomous vehicle to help us understand the challenges that future space traffic management may face. Currently, autonomous vehicles are equipped with numerous sensors and can achieve autonomous driving functions from L3 to L5 levels, and the most basic function is to avoid collisions. Similarly, the number of satellites in space has increased significantly in the past few years, especially with the implementation of the Starlink program, indicating that the number of satellites in the future may increase by an order of magnitude. This is like when cars first appeared, there were few vehicles on the road, and there was no need to consider traffic congestion and collision risks. But as the number of vehicles increases, traffic management and rule-making become crucial.

For satellites, future space traffic management will also face similar challenges, including how to formulate rules to avoid collisions. Satellites also need to have functions similar to autonomous vehicles, such as collision warning and autonomous avoidance. For example, if two satellites have a risk of orbital intersection, they can adjust the orbital angle to avoid a collision. In the future, each satellite may have such a function, which requires the empowerment of the satellite's autonomous control and orbit adjustment capabilities by AI technology. This is similar to applying the AI model algorithm of autonomous vehicles to unmanned aircraft control technology and further extending it to the space field.

At present, although satellites are invisible and intangible in the sky, there are still specialized teams that control the satellites through ground stations to receive telemetry parameters. If the satellite descends in orbit, instructions can be sent to raise it. Many automation strategies have been implemented to reduce the need for manual observation, processing, and instruction sending. The future development direction is to further integrate these strategies into the satellite to achieve autonomous measurement and control, and ultimately achieve unmanned management. This means that the satellite will be able to perform tasks and adjust the orbit autonomously without ground intervention, greatly improving efficiency and safety. With the development of AI technology, we are expected to see the autonomy and intelligence level of satellites in space continuously improve to cope with the increasing space traffic challenges.

Tian Feng: The concepts of Sentinel satellites and Cleaner satellites are very interesting. Cleaner satellites refer to those spacecraft specifically designed to handle space debris or repair faulty satellites. When a satellite enters a faulty state and needs repair, or has been retired, if they are equipped with AI, they can automatically enter the graveyard orbit. The graveyard orbit is a specific area in space used to store satellites that are no longer in use to reduce the risk of collision with other orbiting satellites.

The intelligence mentioned by Mr. Niu is reflected in that the satellite can self-learn and improve through AI. Reinforcement learning in AI is a key concept that allows the satellite to learn through interaction with the environment, rather than just relying on human guidance. This is similar to AlphaZero, which learns the chess game through self-play. When the satellite fails in performing a specific task, these failure data are very valuable and can be used for self-improvement on the terminal. If the computing power is sufficient, the satellite can become more intelligent through self-reinforcement learning.

In addition, if a satellite learns a new skill, it can broadcast this skill to other satellites. In this way, the entire satellite constellation can share this knowledge, making the intelligent upgrade speed of the entire network much faster than the individual intelligent upgrade of a single satellite. This is similar to the situation of autonomous vehicles. For example, if one of the 3 million Tesla vehicles learns how to drive on a dangerous road section, other vehicles can also quickly learn. This group intelligence and self-learning ability of artificial intelligence is one of its most powerful features. In this way, the satellite constellation can continuously improve itself, improve efficiency and safety, and reduce the dependence on ground control.

36Kr: The data collected by satellites in space is huge and complex. After the addition of AI, will there be any differences in the data acquisition of satellites?

Wei Wenyi: When acquiring data, the satellite can perform preliminary data processing, such as data compression, so that the raw data initially collected can be roughly processed on the satellite and then transmitted back to the ground in the form of rough processed products. On the ground, these preliminarily processed data can be used in various application scenarios, bringing many innovations. For example, in the agricultural field, previously, each region had its own algorithm to process remote sensing data, and the same is true for the forestry and other fields such as ecological and environmental protection, because each region has its unique regional characteristics. But now, if all these fields can use a unified data preprocessing method, then a more standardized and unified conclusion can be obtained.