GEN_AI Project: Applied case studies of Machine Learning C-D5041E

The task

The goal of this project is to fine-tune a generative model capable of creating a sequence of actions for a robot to complete a given task.
Specifically, the model should receive the following inputs:

• A list of objects on the table along with their spatial coordinates (x, y, z)
• The task to be performed Based on these inputs, the model should be able to generate the correct sequence of actions to ensure the successful completion of the specified task.

The schema

EOF coordinates

This element represents the robot's position relative to its own reference system.

Objects coordinates

This element represents the position of all objects in the scene, indicating their location relative to the robot's base.
You can select a minimum of 4 objects and a maximum of 8, at your discretion.

User's task

This element represents the task that the user wants the robot to perform.
The model must be able to execute at least two types of tasks of your choice.
If you're lacking inspiration, a list of possible tasks for the robot to perform is provided below.

• pick obj_A and place to obj_B;
• lift obj_A;
• move obj_A forward;
• move obj_A backward;
• move obj_A on the left;
• move obj_A on the right;

Hint for the training: axolotl

It is possible to fine-tune a generative model using various libraries; one of them is Axolotl.
More details about the fine-tuning here.
advice: if you choose to use this library use the docker version of it.

Foundamental steps

1. Analyze the problem, decide on the objects to use, and define the tasks (at least 2)
2. Study how to perform fine-tuning of a generative model
3. Choose the generative model on which to perform fine-tuning
4. Create the dataset
5. Perform the fine-tuning
6. Test and evaluate the model

Expected output

It is necessary to provide a document that explains in detail all the operations performed during the project, also listing:

• The problems encountered;
• Their respective solutions.

It is necessary to evaluate:

• The model's performance;
• Its reliability;
• The inference time.

Please upload the fine-tuned model on HuggingFace.
optional: animation in Matplotlib of the actions performed by the robot

Useful imports

pip  install trl peft torch datasets transformers jupyterlab accelerate tiktoken matplotlib bitsandbytes evaluate scikit-learn