Human-In-The-Loop-Project with MLN

This task is non-prehensile manipullation task, Inspired by "Non-Prehensile Manipulation in Clutter with Human-In-The-Loop" (ICRA 2020)\ The details about this paper is in Paper.

I divided global task planner with "Mixture of Logit Networks" and heuristic local planner. Mixture of Logit Networks (i.e., MLN) can acheive both epistemic uncertanty and aleatoric uncertainty which makes robots to stop the its movement and ask for human help in high uncertainty regions. I implemented MLN from sjchoi86.

For more details, the presentation url is here. Presentation Video

Results

Toy Task

Main Task

Requirements

• pybullet
• pytorch
• attrdict
• matplotlib
• opencv-python

Train humman in the loop

python main.py

Test Heuristic Planner

python heurstic.py