“You act as an autonomous agent that controls the tracking spacecraft.”
This is the first rapid researcher used to see how well ChatGpt can pilot the spacecraft. To their surprise, the large-scale language model (LLM) worked brilliantly, taking second place in the autonomous spaceship simulation competition.
Researchers have long been interested in developing autonomous systems for satellite control and spacecraft navigation. In the future, there will be too many satellites for humans to manually control them. And in the case of deep sea exploration, the limit on the speed of light means that the spacecraft cannot be directly controlled in real time.
You might like it
If you really want to expand the space, you have to let the robot make decisions for yourself.
To encourage innovation, in recent years, aeronautics researchers have created the Kerbal Space Program Differyial Game Challenge, a kind of playground based on the popular Kerbal Space program video games, allowing communities to design, experiment and test autonomous systems in (somewhat) realistic environments. The challenge consists of several scenarios, including the mission of pursuing and intercepting satellites and the mission to avoid detection.
In a paper published in the Journal of Advances in Space Research, an international team of researchers described candidates who are commercially available LLMs, such as ChatGpt and Llama.
Researchers have decided to use LLM as traditional approaches to developing autonomous systems require many cycles of training, feedback and refinement. However, the nature of the Kelval Challenge is to be as realistic as possible. This means a mission that lasts for several hours. This means that continuous improvements to the model is unrealistic.
However, LLM is already trained with a huge amount of text from human writing, which makes it extremely powerful. Therefore, in the best scenario, only a small amount of careful and quick engineering is required.
But can such a model actually pilot a spacecraft?
Related: AI models lie to you to achieve your goals – and it’s not necessary much
Researchers have developed a method for translating a given state and its goals in text form. They then handed it over to the LLM and sought recommendations on how to direct and pilot the spacecraft. The researchers then developed a translation layer that translates the LLM text-based output into functional code that can operate the simulated vehicle.
With small series prompts and some tweaks, the researchers got Chat Gupto and completed many tests of the challenge. (The first place went to the model based on various equations, according to the paper).
And all of this was done before the release of version 4, the latest model of ChatGPT. There is still a lot of work to do, especially when it comes to avoiding “hastisation” (unwanted, meaningless output). This is especially tragic in a real scenario. However, it demonstrates the power that even ready-made LLMs can function in unexpected ways after digesting vast amounts of human knowledge.
Source link
