Statistical Filtering and Control for AI and Robotics Plan monitoring and applications Alessandro Farinelli
Outline • Applications – Water monitoring • Information gathering in MRS – DCOP formulation and heuristic approach • Recognizing relevant situations – Activity recognition for water drones • Plan monitoring – Petri Nets for plan monitoring – Interacting with the humans
Applications for Mobile Robots Logistics (Kiwa System)
Water Monitoring Water as a natural capital, significant effort to ensure quality.
The Water Framework Directive • EU Member States must: – ensure surface and groundwaters achieve ‘good’ status by 2027 – assess (monitor) the quality of surface and groundwater and measures to maintain and improve their status • The Directive seeks to engage all stakeholders through River Basin Planning Process – Communities become Aware, Interested, Enabled.
Current ‘state of the art’
Problems • Main focus precision for Reporting Compliance for Ministers – This requires large numbers of samples to deliver ‘certainties’ – This monitoring is very costly ( €10’s millions), because: – Sampling, transport, laboratory analysis (precision and quality) • Data is always historic (many days before results come), – Little effort on Investigations and Surveillance monitoring and so fail to assess effectively ‘diffuse’ pollution - urban and rural • Too much reliance on statutory bodies to do all the work and their budgets are reducing – Presumed low / uncertain data quality from 3rd party organisations are a barrier to use – citizens not empowered, other available data not used
Outcomes needed for 2020-2050 • Deliver WFD in a more flexible way increasing efficiency • More targeted ‘monitoring’ that identifies sources & enables effective intervention • Appropriate treatment of CSO, run off, multiple / small pollution sources, delivering improvements • Managing and communicating data and outcomes • Engaging society – Reduce barriers to new stakeholders to take up local monitoring and treatment, and data analysis – society value – Harness citizen science capability and influence local situation
INTCATCH: the global picture Biological analyses Chemical analyses Environmental System (WP2) temperature Smart boats Rainfall and stormwater (WP2; 5; 4; 9) wind Appropriate treatment Actions Monitoring (WP2; 3; 8) WAIS (WP7) Decision rules and directives Numerical models DSS Costs and benefits Decision (WP9; 11) Socio – economic system (WP2; 6) drivers Stakeholders and communities Managing authorities Management and control (WP2; 4; 7; 8 10);
Why robotic boats • Sensors Deployed right place right time: effective decision making and management of local ‘diffuse’ pollution EC day2 EC day 1 • Data captured by local stakeholders Citizen Science
Robotic boats for water monitoring ARC boats NUSwan HydroNet Platypus • Large, expensive small, low-cost • Community engagement
Autonomous boats in water propellers airboat • Low-cost • Autonomous • Long-endurance • Easy to deploy
Why autonomy Go beyond line of sight boat Non expert users, ensure good data quality 13
System architecture • The boat can be controlled by a wi-fi connected tablet or a radio controller • The user can define a path on the tablet that the boat follows, navigating autonomously • Different sensors to measure electrical conductivity, temperature, and dissolved oxygen
User interface and path creation The tablet app generates a spiral path to collect data in the area
Data visualization: map overlay Dense geo-localized data for the different parameters
Research issues involved • Information gathering – Where to go next to optimize data collection, coordination • Navigation – Improve autonomous control, recognize relevant situations – Perception, detect obstacle on water • Mission level control – Plan-specify high level actions, monitoring plan execution • Human robot interaction – User friendly interfaces, Interaction with autonomy • …
Coordination for info. gathering Goal : estimate a spatial phenomena with minimum uncertainty Uncertainty contours Mobile sensors with limited sensing and communication range Joint work with R. Stranders, A. Rogers,N. Jennings.
Information Gathering: model • Monitor a spatial phenomena • Model: scalar field – Two spatial dimensions – One temporal dimension
Modeling the spatial phenomena • Use Gaussian Process to model the scalar field • GP powerful model for regression Spatial Correlations Weak Strong
Rewarding information gathering Measure of uncertainty • The value of a sample is 8 connected to the 7 uncertainty reduction 6 • Gaussian Process gives 5 prediction and 4 confidence interval 3 • Use entropy to define Prediction 2 the reward Confidence Interval 1 Collected Sample 0 0 2 4 6 8 10
DCOPs for coordinaton DCOPs: mathematical framework to represent decentralized coordination Why DCOPs ? • Well defined framework – Clear formulation that captures most important aspects – Many solution techniques • Optimal: ABT, ADOPT, DPOP, ... • Approximate: DSA, MGM, Max-Sum, ... • Solution techniques can handle large problems – compared for example to sequential dec. Making (MDP, POMDP)
DCOP Formalisation Agents Variables Variable domains Functions
DCOP: assumptions and objective • Assumptions: – Each variable Xi is controlled by exactly one agent Ai – The agent Ai is responsible for assigning values to the variables it controls – An agent can potentially own more than one variable • Objective: – find the variable assignment such that the sum of all functions is maximised * arg max ( ) x F i x i
DCOP for Mobile Sensors: variables, domains x 2 x x 6 4 x 8 x x 1 7 x x 3 5 Variables Encode Movement
DCOP for Mobile Sensors: utility functions U 2 U U 6 4 U 8 U 1 U 7 U U 5 3 ( ) ( | , ) U x H x x x 3 3 3 1 2 x { , } x x x Utility Functions 3 1 2 , 3 (encode information value) Local Interaction
DCOP for Mobile Sensors: global objective n arg max ( ) U x Fixed movement order i i x 1 i x ,..., x 1 n
DCOP Solution techniques Optimal • No general guarantees on Techniques convergence and optimality ADOPT, OptAPO, • Very good approximation DPOP on practical problems – Affinity Propagation – Survey Propagation Max-Sum Approximated Techniques DSA, MGM Solution Quality
Max-Sum: basic ideas ( | ) H X 2 X 1 x x x x 1 2 2 ( ) H X 1 1 x x 3 3 ( | , ) H X X X 3 1 2 ( ) ( ) z x r x i i k i i ( ) k adj i • Iterative message exchange to build local functions that depend only on agent’s variable • Choose action that maximises local function (z-function)
Max-Sum: messages I ( | ) H X 2 X 1 x x x x 2 1 2 ( ) H X 1 1 x x 3 3 ( | , ) H X X X 3 1 2 sum up info from other nodes local maximization step
Max-Sum messages II From variable i to function j ( ) ( ) q x r x i j i k i i ( ) \ k adj i j From function j to variable i ( ) max ( ) ( ) r x U q x x j i i j j k j k \ i x j ( ) \ k adj j i
Max-Sum on acyclic graphs x ( , , ) F X X X ( ) 2 F 1 X 2 1 2 3 x 1 1 ( ) q x F x x 1 1 1 1 1 3 ( ) max ( , , ) ( ) ( ) q x F x x x r x r x 2 1 1 2 1 2 3 2 2 2 3 2 3 , x x 2 3 • Max-sum Optimal on acyclic graphs – Different branches are independent – q messages remove other variables by maximization – Each variable can build a correct estimation of its contribution to the global problem (z functions)
Max-Sum on cyclic graphs x ( , , ) ( , ) F X X X F X X ( ) 2 F 1 X 2 1 2 3 3 2 3 x 1 1 ( ) q x F x x 1 1 1 1 1 3 ( ) max ( , , ) ( ) ( ) q x F x x x r x r x 2 1 1 2 1 2 3 2 2 2 3 2 3 , x x 2 3 • Max-sum on cyclic graphs – Same computation, but Different branches are NOT independent – Agents can still build an (incorrect) estimation of their contribution to the global problem – Extensive evidence that it works very well in practice
Mobile Sensor Demo • Actions – Move to acquire data in a given location • Single agent utility – Entropy given other agent actions • Global Utility – Sum of conditional entropy values
Situation recognition for Water drones Joint work with A. Castellini, G. Beltrame, M. Bicego, D. Bloisi, J. Blum
Datasets ● 3 datasets ● River/lake ● Sampling interval: 1 sec ● 13 features time latitude longitude altitude Matrix of raw data ( RAW ) NORM speed (10 features) electrical conductivity UNORM dissolved oxygen temperature battery Matrix of processed data ( PRO ) NORM voltage (20 features: mean/std heading sliding window of 10 sec) UNORM acceleration command to propeller 1 command to propeller 2
Recommend
More recommend
