Robotic Arm Trajectory Planner

This project implements a neural network-based planner for a simplified 1-Degree-of-Freedom (DoF) robotic arm (elbow joint), designed to generate smooth trajectories and make task-related choices based on visual input.

Project Overview

The core of this project is a Planner module, implemented using PyTorch, which takes a 2D image of the environment as input and produces:

1. Elbow Joint Trajectory: A smooth, sigmoidal trajectory for the elbow joint over a predefined duration.
2. Multiclass Choice: A "left" or "right" decision based on the color of a target object in the environment.

This simplified version focuses on movements from a 90° starting position to either a 20° (extension) or 140° (flexion) target angle, based on the color of a blue or red target ball.

Currently, there are two planners implemented:

• ANNPlanner: An ANN-based planner that uses a CNN + MLP architecture to process the input image and generate the trajectory and decision. This serves as a baseline for the GLEPlanner.
• GLEPlanner: A planner with similar architecture than the ANN planner taht uses GLE dynamics and learning to generate the trajectory and make decisions.

How to use

Inside the Docker Container

The code in this repository is supposed to be run from within near-nes/controller docker container's project root /sim/controller via:

python -m submodules.pfc_planner.src.train --model gle

To test the pretrained and saved GLEPlanner model, run:

python -m submodules.pfc_planner.src.evaluate --model gle

Locally

The code can also be run outside of the docker container using a fallback implementation for the minjerk trajectory generator.

Install requirements locally in a virtual environment and activate it:

uv venv
uv pip install -r requirements.txt
source .venv/bin/activate

Then run train the GLEPlanner outside of the docker container with:

python -m src.train --model gle

To test the pretrained and saved GLEPlanner model locally, run:

python -m src.evaluate --model gle