Benjamin Murdoch, Matt Ricks, Ales Ball, Jacob Olson, Ethan Durrant, Connor Monson, Lars Watts, and Mark Killpack, Department of Mechanical Engineering
Hod Lipson once asked an audience at TED, “Where are the robots?” We’ve been waiting for
them a long time now and modern robotics still hasn’t caught up with our hopes and
dreams. Through our research we hope to address one of the major topics in robotics: movement
(or kinematics), and use this moment of exploration to build a new low cost robotic hand capable
of use within the exciting realm of aquatics.
Kinematics is broken into two separate fields: forward kinematics and inverse
kinematics. Forward kinematics is a simple problem where the roboticist tracks the angles of
each joint in, for example, a robotic arm, and answers the question, “Where am I now in relation
to where I started? Where in 3D space has my hand ended up after moving a specific joint in my
arm?” In essence, relative position based on geometry, without respect to physics, in order to
track angles along a robotic finger or arm. Inverse kinematics is a much more interesting
problem. It asks, “What do my joint angles need to be in order to reach some given point in
space?” This is the precise question we hope to answer through our research even when
implemented on low cost hardware.
Problems with structures that rely on kinematics include cost and reliability. For example, the
shadow hand, a state of the art robotic hand, can cost as much as $119,7001. The price rises
because kinematic systems are blind to their environment and demand structural consistency.
What if a robotic finger only reaches 33 degrees 87% of the time, or what if tendons used for
actuation slacken over time? Any change in the environment leads to a complete systems failure
for a rigid kinematic system.
Our goal was to create a low cost submersible robotic hand. Unfortunately, with low cost
robotics equipment it introduces a high level of variability, which lead us to need a more precise
controls algorithm. We attempted to handle kinematics and variability through deep learning. For
all those people without the resources to purchase an expensive shadow hand, we used cost
effective 3D printed parts driven by waterproof tendons and actuated by servo motors. Servo
motors were operated via an arduino uno and a 15 channel i2c servo driver. We also had plans to
build sensor grids for each fingertip but were unable to do so within the time allotted for our
project.
Our hardware design simplifies computation and only requires that we have two individual
kinematic mappings for the entire hand. In order to learn these mappings using a neural network
we needed a significant amount of training data. To build our training set we placed a single
finger of the hand into our advisor’s motion capture system and recorded absolute position of tracking dots placed on our finger. Motion capture systems are very commonly used in robotic
control when a system does not yield mathematically consistent feedback.
Our training data was recorded as tuples, which held the absolute position of eight motion
capture dots placed in two centroidal groups on both the tip of the finger and on the palm. With
our tracking dots we captured the absolute position of both the tip of the finger and the palm of
the hand. Data was recorded iteratively: meaning we first captured the position of the finger and
palm at rest, and then selected a random servo point between 0 and 255. After a few seconds of
time we recaptured the absolute position of the finger and recorded the data into observation
tuples.
With our training set constructed of nearly 1000 iterations, we created a four-layer feed forward
neural network using Tensorflow. Neural networks are function approximators loosely inspired
by the biology of the brain. Due to the low number of variables, a small network was chosen
with a cost function designed to optimize the network’s ability to predict the correct servo
reading based on any given set of absolute coordinates given within the same motion capture
environment.
Possible Future work:
In order to make our project more useful and extensible, we will need to begin to shift off of the
absolute coordinate system and change to a relative coordinate system. It would also be useful to
build an arm and wrist, and to finish construction of the remaining fingers to create a fully
functional robotic arm with full range of motion. With the small team we had, we were able to
build a fully functional prototype finger, but were unsuccessful in finishing the construction of
the full hand. If we were to finish construction of the full hand and arm it would be insightful to
use a reinforcement learning algorithm similar to Google Brain’s paper published in 20162
except with a hybrid approach using both neural networks to chose the correct servo setting and
the kinematic mapping.
Our Hand:
1 http://www.androidworld.com/prod76.htm
2 Shixiang Gu; Deep Reinforcement Learning for Robotic Manipulation with Asynchronous Off-
Policy Updates; arXiv:1610.00633