This repository implements a general data-driven framework for robotic-collision detection.

A neural network was designed to regress future states given a robot's current state and action. Stochatsic forward passes are used at inference time in order to produce a belief distribution over the next state of the robot. An exponentially smoothed norm is used between the set of regressed future states and the ground truth collected.

Train deep dynamics neural networks from data collected

  python train_dd.py

Visually evaluate the performance of trained deep dynamics models

  python evaluate_dd_model.py

Generates state actions pairs for training deep dynamics neural networks

  python mtr_bab_coll_data.py

The following V-REP scene needs to be running for mtr_bab_coll_data.py

  dd_motor_babbling.ttt

The following V-REP scene needs to be running for evaluate_dd_model.py

  dd_current_scene.ttt

Update config.ini BASE_DIR with the absolute path to current directory.
Packages needed to run the code:

• numpy
• scipy
• python3
• pytorch
• vrep (vrep has instructions for importing API functions)