M. Asadpour, D. Giustiniano, K. Hummel, S. Hemilicher, S. Egli. Presented by Fabián E. Bustamante

Transcription

1 M. Asadpour, D. Giustiniano, K. Hummel, S. Hemilicher, S. Egli Presented by

2 Search and rescue the days after Japan Earthquake and Tsunami, 11 *A. Adams et al., Search is a time-critical event: When search and rescue mission may become futile, Wilderness & Environmental Medicine, 18(2), 7 Yushu Earthquake, 1 2

3 The vision UAVs to the rescue! Unmanned aerial vehicles (UAV) for data collection and transport in SAR Collect from users or sensors Transmit to other UAVs and/or control station! Advantages and challenges Partially controlled mobility they are robots after all Communication on the move Limited battery autonomy max 3 3

4 Deciding when to communicate! Reduce the problem to 2 UAVs one with collected data (M data ) to deliver Key question send ASAP or wait until it is closer? Send as soon as possible for minimum delay Send later for best throughput and risk loosing data in a failing UAV UAV1 UAV2 d d 4

5 Higher throughput but less time! Three strategies and experimentally measured results Target UAV hovering, the other Transmitted data (MB) d= d= d=6 d=8 moving Start immediately Transmit while moving Move to d and transmit So, should you wait for better results? Time (s) Figure 1: Experimental measurements of transmitted data. 5

6 A model of delayed gratification! Consider only minimizing delay of delivery and chances of failure Data to deliver M data d d Node failure! T tx % of M data T ship T tx 7% of M data T ship T tx % of M data 6

7 A model of delayed gratification! Model as a utility function U(d) U(d) = instantaneous utility (u(d)) * reward discounting (δ(d)) Instantaneous utility 1/Communication delay C delay (d) = T ship + T tx T tx = M data / Throughput at distance d (s(d)) Reward discounting chances of failing as a function of time Assumption failure probability is exponentially distributed with the distance travelled (but failure rate is distance independent)! Key element how bandwidth changes with distance No clue! Let s measure it 7

8 Flying platforms for experimental work! Planes and copters Swinglet a single, electric propeller Arducopter an electric quad-copter Airplane Copter Hovering No Yes Size 8cm wingspan 64x64cm frame Weight 5g 1.7kg Battery autonomy 3 Cruise speed 1 m/s 4.5 m/s Max alt 3m 1m 8

9 Experimental setting!!! Two experiments with two UAVs each Using 82.11n (on 5GHz) for air traffic XBeePro for control channel between ground station and UAVs 2.4GHz band, low bandwidth (<25kbps) and long range (<1.5km) 82.11n XBeePro

10 munication. However, we expect higher aerial throughput than the Throughput withthanks airplanes 82.11g-like observed, to 82.11n features such as channel bonding, A-MPDU frame aggregation, and block ACK. Next, we potential causes.! discuss Measured using UDP traffic and iperf Fixed PHY Rate. We now study the impact of the rate adap! Distance using Haversine formula to GPS tation scheme on the throughput by fixing the PHY rate and using coordinates Median, first and third quartiles; wiskers are max and min (?) Throughput (Mb/s) Seems like 82.11g 3 Auto-rate adaptation in motion seems to be the problem Best modulation and coding scheme changes with distance Internet experimentation the (network) edge Figure 5: Throughput vs. Pushing distance between twoto airplanes. 1

11 6 Throughput (Mb/s) Throughput (Mb/s) ityairplanes than in the airplanes tests shown in Figure 5 (see than in the tests shown in Figure 5 (see the upper a comm-mimo ity lower quartiles of as thewell boxplot wellof asthe thewhiskers ends of lower of the boxplot as theasends ghput than the quartiles e Throughput with quadcopters Impact Speed. of Relative Speed. We takeofadvantage of Impact of Relative We take advantage the high lev ch as channel el control of quadrocopters to analyze howaffects the spt of we control ofof quadrocopters to analyze how the speed CK. e! Next, Higher, more stable throughput while hovering! Worst while moving (at 8m/s) p-he rate adaphovering Moving Hovering Moving d=6 m gate and using v (m/s) Figure 7: Quadrocopters tests. an Figure 7: Quadrocopters tests. On the lefton andthe theleft cent throughput versus distance while hovering m throughput versus distance while hovering and moving,and respe 11

12 We observe higher throughput and smaller variabilplanes tests shown inwith Figurequadcopters 5 (see the upper and Throughput f the boxplot as well as the ends of the whiskers). ative Speed. We take advantage of the high level! The impact of speed adrocopters to analyze how the speed affects the Remember auto-pilot speed 4.5m/s! Should we extend the model to account for this? g Moving d=6 m v (m/s) 15 drocopters tests. On the left Pushing and Internet the center, experimentation to the (network) edge 12

13 Delayed gratification and failure rates! Matlab simulation with seemingly logical parameters and s(d) (throughout at distance) derived from measurements.25 Airplanes.4 Quadrocopters U(d) Nothing surprising optimal.3 distance increases with failure rate 1 3 ρ=.111 ρ=.1 ρ=.2 ρ=.5 ρ=.1 Maximum ρ=.246 ρ=.1 ρ=.2 ρ=.5 ρ=.1 Maximum ( ) 13

14 Summary and some discussion points! In a UAV network it seems better to wait, move first and hover for communication! A holistic view that includes movement for mission & communication is future work! Can the auto-rate algorithm be fixed?! What s the impact of communication on power consumption/failure rate and, thus, the model? 14