SAGA Drone Swarms in the Field Vito Trianni - - PowerPoint PPT Presentation

SAGA
 Drone Swarms in the Field

Vito Trianni
 vito.trianni@istc.cnr.it Workshop on Small UAVs for Precision Agriculture May the 13th, 2018
 Villa Salvati, Pianello Vallesina,
 Monte Roberto, Ancona, Italy

http://laral.istc.cnr.it/saga

Why swarms for PA?

• Parallelise operations ➝ higher efficiency
• Collaborative monitoring ➝ higher accuracy
• Redundant systems ➝ higher robustness
• Decentralised algorithms ➝ higher scalability (group/farm size)

The SAGA project

Hardware Enhancement Onboard
 Vision Swarm-level Control

UAV hardware

• UAV based on the Avular Curiosity platform
• payload of 1kg, 10’ flight time
• RTK-GPS, double IMU
• double real time control cores (Cortex M4F)
• Enhanced for swarm operations
• UWB positioning and communication
• 2.4GHz XBee radio link
• Raspberry Pi RGB camera for onboard vision


and high-level control http://www.avular.com

Onboard weed recognition

altitude: 10m altitude: 3m

Onboard weed recognition

altitude: 10m altitude: 3m

SAGA in a nutshell

Hardware enables:

• communication among UAVs
• high-level control and 

• nboard vision

Onboard vision enables:

• low-altitude weed classification
• high-altitude density estimation

Swarm-level control:

• collaborative weed mapping
• decentralised UAV deployment

Collaborative Weed Mapping

Collaborative Weed Mapping

• Full coverage of a cultivated field to inspect for weeds
• Collaboratively map weed presence minimising classification errors
• Aim at robustness, efficiency and scalability
• Deal with environmental heterogeneities
• Proposed solution: reinforced random walks (RRW)
• Comparison with optimal ‘sweeping’ strategy

Albani, D., Nardi, D., & Trianni, V. (2017). Field Coverage and Weed Mapping by UAV Swarms
 To be presented at the 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2017), Vancouver, Canada, Sept. 2017

RRW for Field Coverage

• The field is divided into cells to be visited
• Agents perform a correlated random walk
• Random selection among those cells


that are closer and not yet visited

• Preferential choice of forward semi-plane
• Persistence controlled by parameter
• Neighbour agents repel each other
• Repulsion controlled by parameter
• Provision of an additional directional bias

p ∈ [0, 1[ σa ∈ [0, 50]

RRW for Field Coverage

• Without repulsion among agents,

high persistence leads to
 lower coverage time

• Small groups are not strongly

affected by repulsion

• The larger the agent density, the

stronger the repulsion

• Persistence is detrimental for large

densities and strong repulsion

RRW for Weed Mapping

• Introduction of communication


with limited range

• Range controlled by


parameter

• Tests performed with varying


re-broadcasting protocols

• Agents can place “beacons” 


to attract other agents

• Attraction controlled by


parameter

• Weed mapping efficiency increases

σb ∈ [4, 32] Rc ∈ [5, ∞[

Decentralised UAV deployment

Decentralised UAV deployment

• Onboard vision and autonomous control

allow for non-uniform coverage

Decentralised UAV deployment

• Onboard vision and autonomous control

allow for non-uniform coverage

• High-altitude estimation of weed density

Decentralised UAV deployment

• Onboard vision and autonomous control

allow for non-uniform coverage

• High-altitude estimation of weed density
• Low-altitude collaborative weed mapping

Decentralised UAV deployment

• Onboard vision and autonomous control

allow for non-uniform coverage

• High-altitude estimation of weed density
• Low-altitude collaborative weed mapping
• Attention should be focused only to those

areas that contain weed patches

Decentralised UAV deployment

• Onboard vision and autonomous control

allow for non-uniform coverage

• High-altitude estimation of weed density
• Low-altitude collaborative weed mapping
• Attention should be focused only to those

areas that contain weed patches

• The problem translates to 


utility-dependent UAV deployment

Decentralised UAV deployment

• Onboard vision and autonomous control

allow for non-uniform coverage

• High-altitude estimation of weed density
• Low-altitude collaborative weed mapping
• Attention should be focused only to those

areas that contain weed patches

• The problem translates to 


utility-dependent UAV deployment

• Solution exploits a design-pattern for

decentralised collective decision making

Decentralised UAV deployment

• UAVs explore and estimate the utility of areas during 


high-altitude/low-resolution inspection

• UAVs perform low-altitude/high-resolution inspection only for high-utility areas
• The utility of areas varies through time as a function of the mapping effort
• UAVs get recruited to areas of high utility
• UAVs are inhibited from monitoring areas when
• ther areas of high utility need attention (cross-inhibition)
• there are too many teammates (self-inhibition)

Albani, D., Manoni, T., Nardi, D., & Trianni, V. (2018). Dynamic UAV Swarm Deployment for Non-Uniform Coverage (pp. 1–9). Presented at the AAMAS '18: Proceedings of the 2018 International Conference on Autonomous Agents and Multiagent Systems.

U A B C

uncommitted:

• high-altitude inspection
• estimate area utility

deployed to an area:

• low-altitude mapping
• recruit/inhibit teammates

deployment:

• spontaneous (utility-driven)
• interactive (recruitment)

abandonment:

• spontaneous (task completed)
• interactive (inhibition)

Models and Simulations

• We study a model of area utility dynamics subject to collaborative mapping
• We identify optimal parameterisations for area inspection


depending on UAV collaboration and potential interferences

• We determine the optimal number N★ of agents for efficient monitoring
• We study a coupled model of deployment and utility dynamics
• We translate model prescriptions into a multi-agent implementation
• We introduce the ratio r between interactive and spontaneous transitions,

and study its effects on deployment

Time (s) Time (s)

N★ N★ N★ N★

Summing up

• Collaborative field monitoring and mapping provides
• parallel operation (efficiency) and collaboration (accuracy)
• robustness and scalability: group size can vary in real time
• Decentralised deployment and re-deployment provides
• ability to focus only on areas of high interest
• ability to enforce utility-responsive strategies

Beyond SAGA

• Swarm robotics is a promising approach for the agricultural sector
• Extensive field tests to support the concept
• Determine the legal and economic framework that make


swarm solutions profitable

• Collaborative perception to improve detection accuracy


beyond simple scenarios

• Exploit perception at different time and from different perspectives
• Determine optimal strategies for information foraging to maximise accuracy
• Collaboration between UAVs and ground rovers


forming a heterogeneous swarm

Thanks for your attention