Building the architecture by hand

Context and role

The problem

Using a Transformer for multi-robot path planning requires representing trajectories, not just tokens. The positional encoding and attention structure have to carry the geometry of the problem.

Most tutorials load a pretrained model and fine-tune it. I wanted to understand the pieces by building them.

The goal was learning, not a benchmark result.

Approach

• Built the final architecture using mostly PyTorch components.
• Implemented positional encoding manually in Python, without an external dependency.
• Applied the architecture to multi-robot path planning using trajectory Transformer ideas.
• Stopped after the architecture was built, before a full training and evaluation phase.

What I did not claim

The boundary of a project is part of its evidence. These are the claims this work does not support.

• Not a completed application and not a published result.
• No benchmark numbers or comparison to published path-planning systems.
• Most components use PyTorch; only the positional encoding was written entirely from scratch.
• No claim of a production robotics deployment.

Results

• A working architecture and a practical understanding of attention, positional encoding and sequence modelling.
• The foundation for later architecture-level work, including the custom loss and fine-tuning in MTF-Qwen.
• An honest example of learning by building, presented without inflated results.

Stack

• Python
• PyTorch
• Transformer architecture
• Positional encoding

What this means for you

The useful part of this project is not the artifact. It is the method: define the measurement first, keep the scope honest, and hand over something your team can run. That method is what transfers to a new problem.

If your situation looks similar, the related services below are the place to start. If it does not, send a short description and I will tell you whether I am the right person.