The Neural-Inspired Robot Controller for Hammering was my project with the USC Brain Project led by Dr. Michael Arbib which I worked on as my senior thesis (Download here).
Research Mission: The goal of this research is twofold –
- To understand how humans use a hammer (i.e. the physics of hammering and how the brain controls the hammering task)
- To use that knowledge towards improving tool use control on a robotic hand.
I am currently pursuing both. The bulk of my current work is a mathematical design for a learning controller for swinging the hammer. The design is influenced by principles in the human motor system, such as Fitt’s Law, a speed-accuracy tradeoff in human motion control.
I hope to see some good simulation results in an in-house arm simulator and then prototype on my robotic arm. An actual hammer for the prototype cannot be used as my robotic arm isn’t industrial but I shall have a working physical system ideally.
To get this project off the ground, I needed to explore robotic literature on manipulation and control in addition to literature on the human motor system. Here is a literature review I compiled last summer here.
You can download my full senior thesis (Download here).
UPDATE: In graduate school, I have still been very interested in this problem. I used two class projects during my Masters’ degree to try to extend some of the frameworks and ideas to advance the cause of neural robotics.
First, for my Kinematics, Dynamics, and Controls Project, I worked on robust calibration methods for the Barrett Robotic Arm using Nonlinear optimization algorithms inspired from gradient descent and other concepts from classical neural networks. (See report here). The use of such optimization helps produce powerful non-linear models that may be useful in controlling human-like robots.
Second, for my Machine Learning project, I worked on a Machine Learning method for localizing brain activation of linguistic semantic features in FMRI images. Results were benchmarked on Prof Tom Mitchell’s Thought Recognition data set (See report here, class poster here). Hopefully, these types of Machine Learning methods will help thought-read user intent better in the future.
All original code for the Robotic Arm Project is open source on tandonexperimental. See: https://github.com/prateekt/tandonexperimental. Code is located in the “HammeringSimulation2” folder.
The Robotic Arm Project is supported by a Vice Provost Research Fellowship from USC. Special thanks to the sponsoring entities for funding the robotic arm equipment for this project.