Скачать или смотреть Using the Adaptability of AI for Overcoming Cold Gas Thruster Failure in SpaceX Booster Landings

A while back I made a project where Reinforcement Learning was used in order to land SpaceX booster rockets. How the rocket works, and how the AI was trained is explained here:
   • Landing SpaceX Booster Rockets with Artifi...  

To summarize, in the previous project, the rocket had 4 cold gas thrusters on the top in order to orientate itself (right, left, forward, backwards). Additionally, it could utilize thrust vectoring (10degrees in each direction) for more directional control.

I got an interesting suggestion to completetly disable one of its key components, in this case the cold gas thruster places on the right of the rocket, and see if the AI could overcome this limitation. To increase the challenge, I also limited the thrust vectoring to 5 degrees in each direction.

After ~6 hours of training, starting with the previous model, it seems that the AI was able to handle the situation pretty well. Although interpreting the end result is difficult, it looks like it tries to alter its incoming trajectory so that the 3 remainig gas thrusters are more effective. Some of the "harder" starting positions result in a rocky landing, however when observing the relevant graphs from the training showed that there was still improvement to be learned. Meaning that letting it train for longer, the landing velocity, as well as the distance to the center, would be lower. (Let me know if you want to see the graphs)

Also important to note is that I thought it would be fun to add a negative reward for each timestep, meaning that the rocket ideally wants to land as fast as possible.

Probably goes without saying: this is not a realistic simulation, more or less a fun side project.

I've decided to take a serious attempt at this problem, so I will recreate a far more realistic simulation environment. Grid fins are one of the most important components which are left out so far. In fact, I haven't seen any AI based models taking this into account. Additionally, sensor data is not always reliable, and there are a lot of techniques which are capabable for handling these issues. I am able to pursue this topic for my MSc thesis, so I'll be working on this project for a full year.