MSc Project: Active Inference for Closed-Loop Robot Control

by
  • Graduation Project
  • Delft

Cognitive Robotics Department, TU Delft

Background

Active inference and the free energy principle are framework formulated by K. Friston [1] to explain the behaviour of biological systems, including human intelligence, making it the most promising theory for true AI. Active inference is explored in neuroscience for a unified account of perception and action, and has been proposed as an alternative to reinforcement learning or optimal control approaches to generate optimised behaviour policies [2]. Preliminary results have been obtained applying this framework in robot control [3]. At the Robot Dynamics lab of the TU Delft Department of Cognitive Robotics, research is being conducted to explore the advantages of using the active inference schema for online control of real robots, with preliminary results using active inference in open-loop.

Objective

The project will explore the possibilities of the biologically inspired active inference framework for the control of industrial manipulators. Dynamic control of industrial manipulators present challenges such as solving the inverse model problem or dealing with noise and modelling errors. This project will:

  1. Design a closed-loop control schema based on active inference for the control of an industrial manipulator.
  2. Create a ROS package for the implementation of active inference-based controllers for industrial manipulators.
  3. Analyse the performance of the designed control schema in different robotic tasks.

Approach

The project will build on previous results in the literature and at TU Delft to design and implement a close-loop control based on the active inference schema for an industrial manipulator:

  1. designing experimental robotic setups in simulation and with a real robot arm to test the designed control schema,
  2. design close-loop control schema under the active inference paradigm for relevant robotic applications, e.g. position control, visual servoing, and
  3. compare its performance to standard controllers for industrial manipulators, evaluating amongst others:
    • robustness to exogenous perturbations (e.g. forces applied to the robot during movement)
    • robustness to deviations in the robot model
    • robustness to sensorimotor delays


References
[1] Friston, K., Kilner, J., & Harrison, L. (2006). A free energy principle for the brain. J Physiol Paris., 100 (1-3), 70-87.
[2] K. J. Friston, J. Daunizeau, and S.J. Kiebel. Reinforcement learning or active inference? PLOS ONE, 4(7):1-13, 07 2009.
[3] L. Pio-Lopez, A. Nizard, K. Friston, and G. Pezzulo. Active inference and robot control: a case study. J R Soc Interface, 13(122), Sep 2016.

To apply for this job email your details to c.h.corbato@tudelft.nl