/rlfd-with-obstacle-avoidance

Robot Learning from Human Demonstrations with Unexpected Obstacles during task reproduction. This is achieved using a combination of Motion Planning (BIT*) and Motion Primitives (KMP).

Primary LanguagePython

Robot Learning from Human Demonstrations with Unexpected Obstacles During Task Reproduction.

Motivation

  • Standard motion primitive frameworks assume that the environment in which a robot is trained will be identical to the environment in which it is used.
  • This assumption doesn’t hold true in most real-world situations

Goal

To develop a system that learns to perform tasks through demonstration and autonomously avoid obstacles during task reproduction

Method

  • Learn tasks through demonstration using Kernelized Movement Primitives
  • Use a motion planner to generate a path with minimum possible points
  • Extrapolate KMP to pass through the points generated by motion planner during task reproduction

Directory Details

  • 2D_toy_problem has the complete system implemented for a 2D obstacle avoidance problem built using pygame
  • src has the files that implement KMPs for a Kuka IIWA arm simulated using Pybullet
  • pickled_objects contains trained kmps as pickle files
  • training_data contains the data used to train KMPs

Dependencies

  • Pyrobolearn