An inventor has created an impressive simulation on a YouTube channel called "gonkee," where they teach a 3D model how to simulate flying using neural networks. The project began by importing a highly detailed 3D model of Iron Man into Blender software, which required reducing the number of triangles to reduce the load on the computer. Joints, collisions, and rigid bodies were added to the model to enable it to fly, and reinforcement learning was used to teach the model how to move in the air.
Reinforcement learning involves training an AI model to make decisions based on the desired outcome, and the method used in this project was proximal policy optimization (PPO). Blender provided an inbuilt PPO, which allowed the creator to create actions for their model. The model was trained to achieve three goals: stopping at a point, moving from one point to another, and maintaining momentum. After training for approximately three million timesteps, the model was able to stay at a stationary point, rotating around its axis.
However, when the model was tested with different points, it failed, and Iron Man fell into an infinite void. To address this issue, a new system was implemented that allowed Iron Man to maintain momentum and retry the same point if he failed. After refining the model and training it for another 10 million timesteps, Iron Man was able to reach the changing points.
To add complexity, a miniature 3D world was created with buildings, plants, and bright orbs for Iron Man to fly through. Although the model worked for a few minutes, it eventually fell off the map.
The inventor assumed that Iron Man's armor design was at fault and added a drone, but the results were the same.
The project can be built by anyone who has knowledge of machine learning and blender. The video has described the important parts needed for it works. So, you can check that out too.
