Animal simulated balance in robots
Kyle R. Snipes
University of Illinois at Urbana-Champaign
2007
Motivation
The functionality and role in the real world of robots is limited by their mobility. It’s not difficult to get a robot to move inside a lab or other isolated environments, but in the real world obstacles arise. Autonomy in motion would open up new functions for them to perform. Autonomous robots would be able to handle search and rescue, construction and inspection, and various other examples more effectively. In order to achieve autonomous motion and overcome unpredictable obstacles, a robot needs to be able to self balance.
Many attempts have been made to program robots to balance themselves using predictive algorithms. These algorithms do not provide real-time feedback and cannot account for all obstacles that may arise. Animals and humans alike possess the mechanisms to balance in a non-predictive, but rather reactive way. It is my goal to develop a robot that uses similar neuronal mechanisms as animals to balance.
Many invertebrates sense their orientation with respect to gravity using statocysts, which are hair cells embedded in a gelatinous membrane. When body position changes, the membrane bends the hair cells accordingly which activate sensory neurons. Vertebrates use a similar structure, found in the inner ear. The hair cells are contained in two chambers known as the utricle and saccule. Both are embedded in a gelatinous membrane called the otolith membrane. Due to their differing orientations within the otolith membrane, the utricle is more sensitive to horizontal acceleration while the saccule is more sensitive to vertical acceleration. The acceleration in either direction causes the hair to bend, generating action potentials in associated sensory neurons. The utricle and saccule labyrinth forms three semicircular canals oriented in three planes to detect angular acceleration in three dimensions. The saccule, utricle, and semicircular canals make up what is known as the vestibular apparatus1.
Proprioceptors are another mechanism used to determine body position. A proprioceptor is a sensory receptor within tendons and joints that provides information about relative position and movement of body parts. Muscle spindles are stretch receptors comprised of several muscle fibers. When a muscle stretches, the spindles are elongated and attached sensory neurons fire. Muscle contraction exerts tension on tendons attached to the muscle, which is monitored by the Golgi organ, another type of proprioceptor1.
To implement a balance system in robots, regardless of which type of receptive system used, there are several factors that come into play. Detection, compensation, and inherent stability all need to be taken into account and achieved and various levels. Detection is the ability of the robot to sense outside forces upon it. The information detected needs to be interpreted in a meaningful way so that the robot may later perform an appropriate action. In balance control, this translates into the detection of deviation of e.g. a limb, from the desired position. Once detection is complete, a robot must compensate for the deviation detected. Again, in balance control this translates into returning the limb to the desired position. When a system overcompensates, a cycle of detection and compensation begins again. The inherent stability is the physical properties of the system that influence the amount of deviation detected and compensated for. A balance system that is inherently stable will naturally be able to balance itself with little compensation. On the other hand, an inherently unstable will enter a cycle of constant detection and compensation.
Design
For this project, I have developed a robot that utilizes a balance system modeled after animal proprioception. The physical design of this project is a simple leg made of balsa wood attached at a pivot. At the top of the leg is an accelerometer (DE-ACCM5G) which acts in a manner similar to a proprioceptor as the detection unit of the control system. Connected by strings to the leg are two basic DC motors. Figure 1 depicts the setup.
Figure 1. The mechanical model of the leg hooked to the vestibular system and motors. The leg falls perpendicular to the axle, with the accelerometer reading the angle.
The leg is attached to the axle in a manner that allows it to fall freely in to the right or the left (towards each motor) perpendicular to the axle pivot. The accelerometer attached to the top of the leg reads its angle position with respect to the X axis. This output is then sent to the corresponding neuron which then generates a signal sent to the corresponding motor. Figure 2 depicts this.
Figure 2. Accelerometer outputs voltage dependent on angle of inclination with respect to the X axis. This voltage is offset by +/- 2.5V and sent to the corresponding neuron which then generates a signal to be sent to the corresponding motor.
The motors spin corresponding to the signal received and will spin, tightening the string and pulling the leg towards the spinning motor. The motors are responsible for compensation of the system.
The leg is designed to be inherently unstable; meaning on it will remain in a cycle of detecting a fall and overcompensating. The leg is attached loosely on the axle to minimize friction and maximize instability.
Principles of operation
At 90 degrees (upright) the accelerometer outputs a value of 2.5V. At 0 degrees (horizontal to the left) it outputs 2.8V and at 180 degrees (horizontal to the right) it outputs 2.26V. This output is then offset by +/- 2.5V via a differentiating opamp.
The opamp differentiator (Figure 2 Left OpAmp, Right OpAmp) works by subtracting the – input from the + input. The 2.5V offset then prevents both neurons from firing at the same time, causing both motors to never run at the same time. This is because when the accelerometer is to the right (output > 2.5V) for example, the left opamp (offset – accelerometer output) produces a negative value and the left neuron doesn’t fire. Similarly, when the leg falls to the left (output < 2.5V), the right opamp (accelerometer – offset) produces a negative value, preventing the right neuron to fire.
Figure 3. The electrical circuit model of the sensory input to the vestibular neuron. Voltage output of the accelerometer is being offset by the voltage divider Rdiv (2.5 volts). This difference is amplified by the OpAmp (AD622N). Amplification gain is set by Rgain (2.7 KOhm). The buffer stabilizes the output of the sensory system.
The neurons used are based on the Hodgkin-Huxley model. Figure 4 depicts the neuron circuit diagram.
Figure 4. The electrical circuit model of the vestibular neuron. Cmem gradually charges up. Once it reaches the threshold voltage set by R1, a spike is produced similar to an actual neuron. The inverter is present to invert the pulse sent from the output from high to low and vice versa. The diode D1 is present to control the direction of the current. Rdisch regulates how much current is drawn out of Cmem which is the rate of discharge.
Figure 5 shows the input-out characteristics of the various electrical components of the simulated vestibular system.
A B
Figure 5. Electrical properties of the artificial vestibular system. A. Input-output characteristic of the OpAmp within the context of the circuit on figure 3 (orange squares) and of the neuron (blue circles). B. Input translation from accelerometer to neuron input (blue circles) and output characteristic of the neuron with respect to angle reading from accelerometer (orange squares).
As you can see from Figure 5A, the accelerometer and neuron (with 5V input) follow a linear input-output characteristic. The non-linearity of the opamp, from Figure 3, causes the input of the neuron to become non-linear which in turn produces a non-linear output. What this lead to was the neuron firing faster as the leg fell until it topped out a certain value. For this project, being inherently unstable was not a bad thing. The purpose was not to build a robot that was fully self-balancing, but instead study how to make one. The input-output response of the neuron was what we were investigating. Being inherently unstable allowed us to better observe how the neuron responded to a falling leg.
Photos
A. / B.Figure 6. Photographs of the physical and electrical components. A. Photographic representation of figure 1. B. Photographic representation of figure 4.
Future Work
To complete my research there is more work to be done. I still need to hook two neurons up to the design. We know the design works in both directions (+X and –X) because I tesed that using a single neuron one direction at-a-time, but I did not have two neurons hooked up at once. Also, after testing we saw that the DC motors do in fact work with the circuit when they are connected by means of a classic H-bridge, but have yet to fully hook them up and tweak them.
More work I would like to do is combine projects with others working alongside me. A more realistic neuron was developed as well as synapse. The neuron developed can handle signal modulation whereas mine does not.
References
1. Peter H. Raven, George B. Johnson, Jonathan B. Losos, Susan R. Singer
Biology 2005:974-979.