Cow Herding with Virtual Fences

From DRLWiki

Revision as of 15:20, 19 August 2010 by Carrick
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Building fences to manage the grazing of cattle can cost upwards of $20,000 per kilometer. These fences have additional maintenance costs and do not provide fine grain control over the area in which the cattle are grazing. We have developed a system to provide virtual fencing and control of herds of cattle. All or some of the cows are outfitted with our system which has position and orientation sensors as well as sound and shock systems to provide feedback to the cattle. We can thus create virtual fences which can be easily moved depending on the conditions of the fields. For instance, a week after an isolated rain storm a rancher may want to move the herd to the area which received the rain and now has new growth.

Our system uses a directional sound system to indicate when a cow is approaching the virtual fence. If the animal does not respond to the sound queue a small directional shock is applied. Preliminary experiments show that the cattle learn quickly to respond to the sound and rarely need to be reinforced with the shock. Limits on the total number of sound and shock queues are enforced to prevent stressing the animal. As such, our system still requires perimeter fencing, but less division on the interior. The rancher is able to see in real time the locations of the animals (and in the future heath status) as well as historical data. This improves land management and utilization.

We are also developing algorithms which model and predict the motions and reactions of the cattle. This allows us to deliver the proper and minimal stimulus to get the desired action. Additionally, by identifying leaders of the herd we will only need to place instruments on a small percentage of the total herd.


Contents for This Page

Hardware Description

File:NewMexico1.mpg File:NewMexico2.mpg


We have developed a suite of electronics which sit on the top of the cow's head. The box contains a GPS reciever, 3-axis accelerometer, 3-axis magnetometer, and a temperature sensor. The box also contains electronics for networking with other boxes, and for applying sound and electrical stimuli to the animal. Our system improves on previous work on animal monitoring hardware [Butler et al., 2006, Wark et al., 2007], by mounting the device on top of the animal’s head, as shown in Figure 2, instead of packaging it as a collar. We found the head mounted device improved several aspects of the device’s performance compared to the previous collar mounting: (1) the GPS satellites were more likely to be visible from the top of the head, (2) solar panels on the box were more likely to receive direct sun exposure, (3) networking radio communication was less obstructed by the animal’s body, (4) the animal was less able to deliberately rotate the box, and (5) the box was prevented from being dipped in water or mud and was generally better protected.

Our sensor box is approximately 21.5cm×12.0cm×5.5cm and weighs approximately 1kg. The processor is a 32bit ARM7TDMI cpu (NXP model LPC2148) with 512kB program memory, 40kB RAM, USB, and a 10 bit A/D converter. The device also has 256kB FRAM (external non-volatile memory with no rewrite limit) and a removable SD card with 2GB storage capacity. Data can be easily and quickly downloaded to a computer by physically transferring the SD card, or by downloading remotely via the radios. There are 2 hardware serials which are multiplexed for a total of 5. The sensors in the box include a GPS engine, 3-axis accelerometer, 3-axis magnetic compass, and an ambient air temperature sensor. There are many general purpose analogue and digital I/O lines, so additional sensors can be included.

The communication system consists of two radios. Firstly, a 900MHz radio (Aerocomm AC4790) with 1 watt transmit power is used for long range, low band width communication. This radio has a claimed 32km range and a claimed 57600b/s transfer rate. However, we observed a maximum of only 2km range and a data transfer rate of only 1000b/s. This is particularly odd as the flat, remote environment in which the radios were tested should have been ideal for radio transmission. The cause for the poor performance of this radio is still unknown. Secondly, the box uses a Bluetooth radio with 100m range and 100kb/s data rate for short range, high band width communication.

Power is provided by a bank of 8 Lithium-Ion batteries with a total capacity of 16 watt-hours. The batteries are continuously recharged by a solar panel mounted on the top of the box allowing the box to run indefinitely under normal conditions. The batteries have enough capacity for over a week of operation without the solar panels.

Finally, we have a two-tier animal control system consisting of a set of speakers for applying arbitrary, differential sound stimuli and a set of electrodes that enable the application of differential electrical stimuli.

The box’s operating system is a custom designed collaborative multitasking architecture. Processes run as scheduled events which can be scheduled to run at millisecond intervals with no preemption or real-time constraints. The software supports arbitrary network topologies for communication. Users interact with the system via a serial console or a Java user interface. These can be accessed directly through the serial port or remotely over either of the radios. This allows remote reconfiguration of the monitoring devices in the field. The operating system can be completely reprogrammed using an attached serial cable, remotely over the radio, or by placing a file on the SD card.

Model Description

We are developing the tools for automatically generating dynamical models and control strategies for groups of animals based on recorded tracking data. It is envisioned that the resulting techniques will be useful for modeling and controlling a broad class of biological group phenomena including cow herding, bird flocking, insect swarming, and human crowd behavior. Specific applications are numerous. Notably, these tools would enable the monitoring and control of groups of people to alleviate foot traffic congestion, or to identify anomalous behavior. In addition, such methods could be directly useful for the control of groups of autonomous mobile robots. For this project we seek to control groups of livestock to minimize environmental damage from over-grazing. Methods from nonlinear control, system identification, and statistical learning theory are currently being adapted to these purposes.

We have proposed a simple difference equation model with four important features. Firstly, each animal is given individual dynamics to enforce kinematic constraints and basic physical laws. Secondly, a force-like contribution is applied to each animal from its interaction with each of the other animals in its group. This force law is designed to accommodate a principle common to many multi-body dynamical systems: forces between bodies are repulsive at close distances and attractive at far distances (see Figure 3). Thirdly, a force contribution is applied to each animal from its interaction with the environment. The force is a function of the animal's position, inducing a force-field over the environment. This force-field is parameterized in such a way that closed streamlines (orbits) are possible (see Figure 4).

Figure 3. Agent-Agent interaction force magnitude.
Figure 4. Agent-Environment interaction force-field.

Lastly, unknown elements of an animal's decision-motive processes are modelled as a force-like random disturbance.

Systems Identification with Least Squares Fitting

The model structure discussed above has the convenient property that it is linear in its unknown parameters. We use this fact to perform a least squares regression using measured data from real animals to estimate the unknown parameters. THis gives a dynamical model of the cow herd learned from real cow data. The performance of a learned model using the technique is shown in Figure 5. This work is described in detail in File:SchwagerFSR07Cows.pdf, File:SchwagerJFR08.pdf.

File:CowModelMovie.mpg

We also have used a K-means classification algorithm to categorize stretches of animal data into higher-level states File:SchwagerCEA07Cows.pdf. Each state will then have an associated set of regressed model parameters. We plan to model the transition among these higher-level states as a Markov model and identify the state-transition probabilities from the data using maximum likelihood techniques.

Control

Several strategies exist for the control of multi-body dynamical systems. Control laws depend on what actuation is available and the specific dynamics of the agents and their interactions. Experiments have been carried out on herds of cattle using collars that emit a sound to stimulate the animals' movement in a particular direction. However, developing useful control actuators for groups of animals posses a significant practical challenge. We have explored a new means of low stress herd management using the fact that a stressed individual tends to propagate its stress to neighbors, and tends to move toward the center of the herd. We show in [6] that these two facts alone can be used to steer a herd using simple, adirectional on-off actuators on a subset the animals. What is specifically interesting is that simulations suggest that we can achieve directional control using actuators that do not provide directional cues. This paper won the Best Paper Award at the 2008 Tenth International Conference on the Simulation of Adaptive Behavior (SAB '08).

File:HybridModelMovie.mpg

References

[1]

M. Schwager, D. M. Anderson, Z. Butler, D. Rus - Robust Classification of Animal Tracking Data
Computers and Electronics in Agriculture 56:46-59, March 2007
Pdf Bibtex
Author : M. Schwager, D. M. Anderson, Z. Butler, D. Rus
Title : Robust Classification of Animal Tracking Data
In : Computers and Electronics in Agriculture -
Address :
Date : March 2007
[2] M. Schwager, D. Rus - Data-Driven Identification of Group Dynamics for Motion Prediction and Control
Proceedings of the Conference on Field and Service Robotics , Chamonix, France, July 2007
Pdf Bibtex
Author : M. Schwager, D. Rus
Title : Data-Driven Identification of Group Dynamics for Motion Prediction and Control
In : Proceedings of the Conference on Field and Service Robotics -
Address : Chamonix, France
Date : July 2007
[3] M. Schwager, C. Detweiler, I. Vasilescu, D. M. Anderson, D. Rus - Data-Driven Identification of Group Dynamics for Motion Prediction and Control
Journal of Field Robotics 25(6-7):305-324,2008
Pdf Bibtex
Author : M. Schwager, C. Detweiler, I. Vasilescu, D. M. Anderson, D. Rus
Title : Data-Driven Identification of Group Dynamics for Motion Prediction and Control
In : Journal of Field Robotics -
Address :
Date : 2008
[4] Nikolaus Correll, Mac Schager, Daniela Rus - Social Control of Herd Animals by Integration of Artificially Controlled Congeners
Proc. of the 10th International Conference on Simulation of Adaptive Behavior (SAB). Springer Lecture Notes in Artificial Intelligence LNAI 5040 pp. 437-447, Osaka, Japan, July 2008
Pdf Bibtex
Author : Nikolaus Correll, Mac Schager, Daniela Rus
Title : Social Control of Herd Animals by Integration of Artificially Controlled Congeners
In : Proc. of the 10th International Conference on Simulation of Adaptive Behavior (SAB). Springer Lecture Notes in Artificial Intelligence LNAI 5040 -
Address : Osaka, Japan
Date : July 2008

People

  • Dean Anderson (PI, USDA)
  • Daniela Rus (PI, MIT)
Personal tools