← back to blog

Capstone - Week One

431 words 3 min

Week Update

This week began by researching the best possible route to go with the AI model that will be used for post training throughout this project.

The options that I explored included: LLM

  • this is a commonly used solution throughout the literature that I explored in preparation to this project. Many of the papers that I read used GPT4 for training
  • Because this is commonly used in published papers it may not be the route to go as we want to contribute to new findings
  • this option will likely require the usage of tokens in order to run GPT4 or any other LLM option, which may become expensive over time

SKRL/RSL-rl

  • both SKRL and RSL-rl are two options for open source RL libraries, but usage would be very similar
  • both libraries are supported by Isaac Sim
  • both libraries are fairly simple and modular which may allow for a more simplistic approach to training the simulation, I need to do more research on the usage of these libraries for complex procedures or movements that may be required as well as the methods of training.

Groot pipeline/ Groot N1

  • This option is an Nvidia product, highly compatible with and often used in Isaac Sim
  • Groot has billions of parameters. post training is extremely high and would require a higher amount of computational power than the other options
  • Groot can be trained through a variety of training methods including video.
  • I have found options for robotic arms that are specifically designed for use with Groot in Isaac Sim, a lot of documentation for this pipeline is also available
  • this may be a more scalable option and has availability for training purposes, however this may be overkill for the scope of the current capstone

At this point I feel that it would be best to pursue research on both the Groot pipeline and SKRL, this may need to change as I do more research.

After looking into the computer requirements of Groot I doublechecked the requirements needed to use Isaac Sim. I realized that my laptop does not have the computational power to reliably run Isaac Sim during training. The rest of my week was spent researching and constructing a new desktop computer that is powerful enough to run Isaac sim as well as give better options for what we can do as far as AI training is concerned.

Specs

below is a rundown of the specs of my new machine:

  • Os: CachyOS
  • CPU: Intel Core i9 (14th gen)
  • Cores: 24
  • RAM: 32GB
  • Storage: 2TB + 500GB
  • GPU: Nvidia GeForce RTX 5080
  • Vram: 16GB