Mobile Robots

Our outdoor mobile robot experiments use two modied Segway RMP 400 robots. An image of the robots is shown in Figure 6.2. The rst robot is equipped with a Velodyne 3D Lidar with a 360◦ _{eld of view. The second robot is equipped with a}

2D SICK LMS291 horizontally mounted laser scanner with a 180◦ _{eld of view. Each}

robot is also equipped with a server-class computer with an eight-core processor. For localisation, the two robots rely on high-accuracy Novatel inertial measurement unit (IMU) and dierential global positioning system (DGPS) modules.

Figure 6.3  The outdoor experimental site.

6.2.2 Experiment Site

The outdoor experiments were conducted in a semi-urban environment at The Uni- versity of Sydney. The site is a rectangular-shaped lawn outside the ACFR with approximate dimensions of 12m×30m shown in Figure 6.3.

6.2.3 Ground Station

The ground station used for the outdoor experiments comprises two standard laptop computers, one for each robot, functioning as control stations and monitors and a third laptop functioning as the main control terminal. Communication with the robots takes place over a 5GHz 802.11a WiFi network. The network is formed of two Netgear wireless access points connected using 802.3ab gigabit Ethernet at the ground station and a Netgear wireless network card in each of the robots. The o- board stationary camera used as a third sensor node in the experiments is a Prosilica GC2450 camera acting as a bearing-only sensor located at the ground station.

6.2 Outdoor Experimental System 101

6.2.4 Communication System

The communication system is depicted schematically in Figure 6.4. Each of the two robots connects directly to its own access point over 5GHz WiFi. These two wireless links are set to a separate channels to prevent interference. Experimental testing conducted on the wireless hardware resulted in a maximum data-rate of 20Mbps for each link. The two access points are connected to a gigabit Ethernet switch to which the camera and control station computers are connected. A third access point also connected to the switch, denoted as WAP3 in the gure, provides internet access to the system.

Robot1 Robot2

WAP1 WAP2

WAP3

Control Station Switch

Camera

5GHz WiFi 5GHz WiFi

to internet

Figure 6.4  Communication system used for outdoor experiments. Solid lines represent gigabit Ethernet connections while dashed lines represent wireless connections.

6.2.5 Software

All DIF algorithms are implemented as a distributed multi-node ROS system. An EKF implementation is used for estimation and data fusion. For the experiments, the DIF algorithms are added as an auxiliary layer to an existing LIDAR object detection system [22, 70]. Object detection for camera images is achieved using ducial markers generated and detected by ArUco library [31]. The min-cost-DIF and threshold-DIF experiments rely on the path planner described in [106]. On the other hand, negotiation-DIF experiments use a discrete-action xed-horizon planner

that selects actions based on an exhaustive forward search. For simulations, the DIF algorithms and path planning software are identical to the case of hardware experiments. Sensor observations are simulated at the point observation level. The platform is also simulated through a simplied dynamics model.

6.3 Min-Cost-DIF

This section presents the results of our solution to min-cost-DIF demonstrated on scenarios where the link costs are readily attainable. We present results for two scenarios. The rst scenario was tested in simulation and involves two robots. The purpose of this scenario is to validate the advantage of our solution to min-cost- DIF over simple down-sampling of information. It also highlights the benets of assuming the ability for multicast. The other scenario was tested in both simulation and hardware and involves a robot aided by a processing ground station. The purpose of this scenario is to demonstrate the exibility of min-cost-DIF in allowing for the dynamic selection of a sensor-data processing location.

6.3.1 Two-Robot Simulation

We evaluated our solution to min-cost-DIF for a scenario consisting of two mobile robots tracking a moving target. The aim of this simulation is to demonstrate the multicast behaviour of dynamic information ow and to show the improvement in information gain realised in the min-cost-DIF setting in comparison to uniform down- sampling of data rates.

The experimental setup is shown in Figure 6.5 and is based on the outdoor experimen- tal system described in Section 6.2. The two mobile robots are assigned to separate workspaces with approximate dimensions of 5m×5m each. The rst robot has a 360◦

eld-of-view sensor and the second robot has a 180◦ _{eld-of-view sensor. The sensors}

6.3 Min-Cost-DIF 103

Figure 6.5  Demonstration setting of the two-robot simulation and experiment. Robots 1 and 2 are shown with their boundaries. Sample target tracks are shown making a square pattern inside the region of interest.

tracked by the robots moves in a circular pattern within a square region of interest outside the robots' workspaces. The region is approximately 10m×10m in size. The network diagram of the system is shown in Figure 6.6. The object detection routine on each robot imposes a processing cost. Due to the multicast property, the processing cost of each processing module is distributed among the receiving estimators. Therefore, if the estimators collectively evaluate an observation utility greater than the processing cost, then the sensor raw data should be processed and sent forward. Virtual links, not shown in the gure, allow for the no-send policy. The cost of the links incident onto the estimators is adjusted throughout the simulation by the robots' sensor utility evaluations.

To validate the performance of the algorithm, a control test was also run. In the control test, the sensor rates were reduced to the same average rate used in the dynamic case. The control test is also referred to as down-sampled communication.

Robot 1 Robot 2

Laser Laser

Object

Detection DetectionObject EKF and Path

Planner EKF and PathPlanner

0.4 0.4

Figure 6.6  The min-cost-DIF network diagram for the two-robot simulation. Virtual links are omitted for clarity.

The information for each of the robots' estimate over time is shown in Figure 6.7 and the corresponding average bars are shown in Figure 6.8. In Figure 6.7a, the informa- tion in Robot 1's estimate is higher on average for the dynamic ow case. By using a xed rate, the performance of the down-sampled displayed long periods of poor information gathering performance for particular system congurations; for example, when the target was outside the eld of view of Robot 2. This behaviour was pre- vented in the case of dynamic ow since inter-robot communication was boosted when required. The advantage of the dynamic ow case can also be seen more distinctly in Figure 6.8a. The dynamic case for Robot 2 observed some lag in information gathering performance initially. This may be attributed to the delay in sensor utility estimation. The dynamic case eventually outperformed the down-sampled case as shown in Figure 6.7b and its overall advantage is conrmed in Figure 6.8b.

The ow rates and estimated sensor utilities for the dynamic ow case are plotted against time in Figure 6.9. The ow rates shown correspond to the percentage of observations transmitted while the sensor utilities correspond to the resulting dier- ence in the log determinant of the estimate's covariance matrix after an observation. The plots show consistency between sensor utility and ow. In particular, the oblique parts of the ow curve correspond to time periods where a robot would receive sensor observations freely due to raw data already being processed for the other robot. This free reception of data is a feature of multicast routing.

6.3 Min-Cost-DIF 105 As noted in Section 4.1.3, discrete ow decisions minimise the error from our maxi- mum ow approximation of the total ow in each link. As seen in Figure 6.9, with the exception of transient periods, the ow variables were mainly either 0 or 100.

0 50 100 150 200 250 300 3 4 5 6 7 8 9 10 11 12 13 Time (s)

Information (nats) (offset=10.3)

Downsampled Dynamic (a) Robot 1 0 50 100 150 200 250 300 0 5 10 15 Time (s)

Information (nats) (offset=10.3)

Downsampled Dynamic

(b) Robot 2

Figure 6.7  The information value (negative entropy) of the robots' target estimate for the two-robot simulation. The plots are shown for two communication methods, sub-sampled and dynamic, with both requiring the same amount of computation on average. The values correspond to the negative entropy with the addition of an oset value.

6.3 Min-Cost-DIF 107 Downsampled Dynamic 0 1 2 3 4 5 6 7 8 9 Information (Robot 1)

Information (nats) (offset=10.3)

(a) Robot 1 Downsampled Dynamic 0 1 2 3 4 5 6 7 8 9 Information (Robot 2)

Information (nats) (offset=10.3)

(b) Robot 2

Figure 6.8  Time averages of the plots of Figure 6.7 shown in bar format. The improvement for the dynamic case over sub-sampling is clearly observed.

0 50 100

Flow

Robot 1’s Sensor to Robot 1’s Estimator

0 0.5 1 Utility 0 50 100 Flow

Robot 1’s Sensor to Robot 2

0 5 10 Utility 0 50 100 Flow

Robot 2’s to Robot 2’s Estimator

0 1 2 Utility 0 50 100 150 200 250 300 0 50 100 Flow

Robot 2’s Sensor to Robot 1

Time (s)

0 50 100 150 200 250 3000

10 20

Utility

Figure 6.9  Flow rates and sensor utility for the two-robot simulation. Flow rates are shown in solid lines while the evaluated sensor utility is shown in dashed lines. The ow rates shown correspond to the percentage of observations transmitted while the sensor utilities correspond to the resulting dierence in the log determinant of the estimate's covariance matrix after an observation.

6.3 Min-Cost-DIF 109