for AI and Robotics Plan monitoring and applications Alessandro - - PowerPoint PPT Presentation
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
Alessandro Farinelli
– Water monitoring
– DCOP formulation and heuristic approach
– Activity recognition for water drones
– Petri Nets for plan monitoring – Interacting with the humans
Logistics (Kiwa System)
Water as a natural capital, significant effort to ensure quality.
– 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
Basin Planning Process – Communities become Aware, Interested, Enabled.
– 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)
– Little effort on Investigations and Surveillance monitoring and so fail to assess effectively ‘diffuse’ pollution - urban and rural
budgets are reducing – Presumed low / uncertain data quality from 3rd party
– citizens not empowered, other available data not used
effective intervention
sources, delivering improvements
– 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
Environmental System (WP2)
Rainfall and stormwater temperature wind Smart boats Biological analyses Chemical analyses Decision rules and directives Numerical models DSS Managing authorities Stakeholders and communities
Monitoring (WP2; 3; 8) Decision system (WP2; 6) Management and control (WP2; 4; 7; 8 10);
Socio–economic drivers Costs and benefits (WP9; 11)
WAIS (WP7)
Actions Appropriate treatment (WP2; 5; 4; 9)
making and management of local ‘diffuse’ pollution
EC day2 EC day 1
HydroNet ARC boats NUSwan Platypus
in water propellers airboat
13
Non expert users, ensure good data quality
boat
Go beyond line of sight
wi-fi connected tablet or a radio controller
the tablet that the boat follows, navigating autonomously
electrical conductivity, temperature, and dissolved
The tablet app generates a spiral path to collect data in the area
Dense geo-localized data for the different parameters
– Where to go next to optimize data collection, coordination
– Improve autonomous control, recognize relevant situations – Perception, detect obstacle on water
– Plan-specify high level actions, monitoring plan execution
– User friendly interfaces, Interaction with autonomy
Uncertainty contours
Joint work with R. Stranders, A. Rogers,N. Jennings.
Goal: estimate a spatial phenomena with minimum uncertainty Mobile sensors with limited sensing and communication range
– Two spatial dimensions – One temporal dimension
Weak Strong Spatial Correlations
connected to the uncertainty reduction
prediction and confidence interval
the reward
1 2 3 4 5 6 7 8 2 4 6 8 10
Prediction Confidence Interval Collected Sample
Measure of uncertainty
DCOPs: mathematical framework to represent decentralized coordination Why DCOPs ?
– Clear formulation that captures most important aspects – Many solution techniques
– compared for example to sequential dec. Making (MDP, POMDP)
Agents Variables Variable domains Functions
– find the variable assignment such that the sum of all functions is maximised
– 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
* i i x
1
2
3
4
5
6
7
8
Variables Encode Movement
1
2
3
4
5
6
7
8
Utility Functions (encode information value)
2 1 3 3 3
3 , 2 1 3
Local Interaction
n i i i x
1
Fixed movement order
n
1
Solution Quality
Approximated Techniques DSA, MGM Optimal Techniques ADOPT, OptAPO, DPOP Max-Sum
convergence and
– Affinity Propagation – Survey Propagation
depend only on agent’s variable
1
2
3
) | (
1 2 X
X H ) (
1
X H ) , | (
2 1 3
X X X H
1
2
3
) (
i adj k i i k i i
1
2
3
) | (
1 2 X
X H ) (
1
X H ) , | (
2 1 3
X X X H
1
2
3
sum up info from other nodes local maximization step
j i adj k i i k i j i
\ ) (
i j adj k k j k j j i i i j
j
\ ) ( \
x From variable i to function j From function j to variable i
1
2
3
) (
1 1 X
F ) , , (
3 2 1 2
X X X F
) ( ) ( ) , , ( max ) (
3 2 3 2 2 2 3 2 1 2 , 1 1 2
3 2
x r x r x x x F x q
x x
1 1 1 1 1
) ( x F x q
– 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)
1
2
3
) (
1 1 X
F ) , , (
3 2 1 2
X X X F
) ( ) ( ) , , ( max ) (
3 2 3 2 2 2 3 2 1 2 , 1 1 2
3 2
x r x r x x x F x q
x x
1 1 1 1 1
) ( x F x q
) , (
3 2 3
X X F
– 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
– Move to acquire data in a given location
– Entropy given other agent actions
– Sum of conditional entropy values
Joint work with A. Castellini, G. Beltrame, M. Bicego, D. Bloisi, J. Blum
time latitude longitude altitude speed electrical conductivity dissolved oxygen temperature battery voltage heading acceleration command to propeller 1 command to propeller 2
Matrix of raw data (RAW) (10 features) Matrix of processed data (PRO) (20 features: mean/std sliding window of 10 sec) NORM UNORM NORM UNORM
Upstream/downstream navigation (UDN) Manual/autonomous drive (MAD)
Time intervals manually labeled by experts Information sources:
Partial labeling: uncertain intervals were left unlabeled
Data Labeling
GMM From 2 to 8 clusters 300 re-initializations K-means initialization Training performed by EM algorithm Full covariance matrix Max number of iterations: 100 Inference: max model probability HMM From 2 to 8 clusters Observation model: single component multivariate Gaussian distributions Initial state distribution: uniform Initial transition matrix: random stoc. Initial means: computed by k-means Training performed by EM algorithm Max number of iterations: 20 Inference: Viterbi algorithm K-means From 2 to 8 clusters Euclidean distance 300 re-initializations Affinity propagation Preference parameter from 30 to 180 with step 30 times the value of the median of the similarity matrix
P(C1)=1/6*max(5,1,0)=5/6 P(C2)=1/6*max(1,4,1)=4/6 P(C3)=1/5*max(2,0,3)=3/6
Purity: external measure of the extent to which clusters contain a single class
Silhouette: internal measure that contrasts the average distance to elements in the same cluster with the average distance to elements in other clusters
Purity GMM HMM KM AP GMM HMM KM AP Silhouette UDN Purity GMM HMM KM AP GMM HMM KM AP Silhouette MAD
Model performance distributions
Five dimensions of analysis: 1) GMM, HMM, KM, AP 4) NORM, UNORM 2) ESP2, ESP4, ESP5 5) # clusters 3) RAW, PRO Total: 324 models, 2 activities
Performance of the best purity models for each method (GMM, HMM, KM, AP)
Performance of the best purity models for each method (GMM, HMM, KM, AP)
Segmentation achieved by the best models
downstream navigation (58% and 37% of coverage, respectively) Best matches for upstream navigation (67% and 16% of coverage, respectively) max silhouette Legend
Statistically significant differences between clusters 1 and 3 (Student’s t-test)
Method: K-means Data: PRO and UNORM # clusters: 7
We are working on a system for collision avoidance based on visual data only
several low-cost platforms, few operators team plans
Plan monitoring for Multi-Agent Systems:
PNP:
– Sensing during execution – Time extended actions, concurrency, interrupts – Distributed execution – Used in several robotics applications – Petri Net Plan page
synchronization
Place Token Weight
Joint work with A. Bertolaso, M. Raeissi, R. Muradore
– Ordinary actions – Sensing actions
50
– Fork
51
– Join
52
– Interrupt
53
UGV Controller Sync UAV Controller
Interrupt change plan execution without aborting Boats to visit locations Swap battery back to visit plan Human decides to interrupt we know we can recover (no failures)
Joint work with M. Raeissi, N. Brooks, N. Marchi, P. Scerri
Hierarchical Petri Net
Coloured Petri Nets: tokens have a type (color)
compact approach to represent team behaviors
Only boats that completed the path
Proxy interrupt (Boat pull out) General interrupt (General Alarm)
– Coordinate robots to maximize information – DCOP is a powerful framework – Careful design of the model (e.g., reward) is key
– High level information on drone activity is important but difficult (unsupervised, real-time, …) – Need specific approaches to have on-line, high-level control
– Formal language to represent high level instructions – Petri Net can represent complex interactions – Coloured Petri Net can be significantly more compact