1 of 18

Distributed Sampling-based Planning for Non-Myopic Active Information Gathering

Mariliza Tzes, Yiannis Kantaros and George J. Pappas

IROS 2021, Prague, Czech Republic

2 of 18

Active Information Acquisition

2

3 of 18

Distributed Sampling-based Active Information Gathering - Overview

3

Goal: Design informative paths for a collection of team of robots to reduce uncertainty over a hidden state

Feasible path

Each robot builds its own tree that explores the robot and information space

Every robot holds and updates its own distribution for the hidden state

Explore the robot-motion space through random sampling

Exchange their distributions and

update using Distributed Kalman Filter

4 of 18

Literature Review

4

  • Myopic

Hoffmann, G. M., & Tomlin, C. J. (2009). Mobile sensor network control using mutual information methods and particle filters. IEEE Transactions on Automatic Control55(1), 32-47.

Dames, Philip, et al. "A decentralized control policy for adaptive information gathering in hazardous environments." 2012 IEEE 51st IEEE Conference on Decision and Control (CDC). IEEE, 2012.

  • Non-Myopic

Search-based

Atanasov, Nikolay, et al. "Decentralized active information acquisition: Theory and application to multi-robot SLAM." 2015 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2015.

Sampling-based

Lan, X., & Schwager, M. (2016). Rapidly exploring random cycles: Persistent estimation of spatiotemporal fields with multiple sensing robots. IEEE Transactions on Robotics32(5), 1230-1244.

Kantaros, Yiannis, et al. "Asymptotically Optimal Planning for Non-Myopic Multi-Robot Information Gathering." Robotics: Science and Systems. 2019.

5 of 18

Literature Review

5

  • Myopic

Hoffmann, G. M., & Tomlin, C. J. (2009). Mobile sensor network control using mutual information methods and particle filters. IEEE Transactions on Automatic Control55(1), 32-47.

Dames, Philip, et al. "A decentralized control policy for adaptive information gathering in hazardous environments." 2012 IEEE 51st IEEE Conference on Decision and Control (CDC). IEEE, 2012.

  • Non-Myopic

Search-based

Atanasov, Nikolay, et al. "Decentralized active information acquisition: Theory and application to multi-robot SLAM." 2015 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2015.

Sampling-based

Lan, X., & Schwager, M. (2016). Rapidly exploring random cycles: Persistent estimation of spatiotemporal fields with multiple sensing robots. IEEE Transactions on Robotics32(5), 1230-1244.

Kantaros, Yiannis, et al. "Asymptotically Optimal Planning for Non-Myopic Multi-Robot Information Gathering." Robotics: Science and Systems. 2019.

6 of 18

Literature Review

6

  • Myopic

Hoffmann, G. M., & Tomlin, C. J. (2009). Mobile sensor network control using mutual information methods and particle filters. IEEE Transactions on Automatic Control55(1), 32-47.

Dames, Philip, et al. "A decentralized control policy for adaptive information gathering in hazardous environments." 2012 IEEE 51st IEEE Conference on Decision and Control (CDC). IEEE, 2012.

  • Non-Myopic

Search-based

Atanasov, Nikolay, et al. "Decentralized active information acquisition: Theory and application to multi-robot SLAM." 2015 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2015.

Sampling-based

Lan, X., & Schwager, M. (2016). Rapidly exploring random cycles: Persistent estimation of spatiotemporal fields with multiple sensing robots. IEEE Transactions on Robotics32(5), 1230-1244.

Kantaros, Yiannis, et al. "Asymptotically Optimal Planning for Non-Myopic Multi-Robot Information Gathering." Robotics: Science and Systems. 2019.

7 of 18

Contributions

  1. We introduce the first distributed, sampling-based active information gathering algorithm.

  • The algorithm is proven to be probabilistically complete and asymptotically optimal.

  • The proposed scheme significantly decreases the computational complexity per iteration of its centralized counterpart.

  • The algorithm can handle large-scale estimation tasks (scalable).

7

8 of 18

Problem Formulation

8

Robot Dynamics

Hidden State Dynamics

Observation Model

(LiDAR, stereo-camera, bearing sensor)

(target dynamics, environmental field)

finite set of control inputs

9 of 18

Communication - Distributed Kalman Filter

9

  • Robots communicate under an underlying communication graph

3

2

1

  • Graph is assumed to be time-invariant
  • Exchange with neighbors their beliefs about the hidden state
  • Each robot updates its distribution (a-posteriori distribution) as a geometric mean of its

neighbors' beliefs and its own observation likelihood function

  • Each robot holds a Gaussian prior distribution over the hidden state

10 of 18

Active Information Gathering Formulation

10

Goal: Given initial robot states and a prior distribution over the hidden state ,

compute control inputs and a planning horizon F for the optimal problem

information threshold

obstacle avoidance

robot, hidden-state,

observation dynamics

accumulated mutual information

smallest possible time horizon

11 of 18

Active Information Gathering Formulation

11

Goal: Given initial robot states and a prior distribution over the hidden state ,

compute control inputs and a planning horizon F for the optimal problem

Stochastic Control Problem

Deterministic Optimal Control Problem

1. Linear observation dynamics wrt

the hidden state

2. Linear gaussian hidden-state dynamics

a-posteriori covariance matrix of the hidden state

Distributed Kalman Filter

12 of 18

Distributed Active Information Gathering

12

Basic Idea: Each robot builds its own local tree that explores both physical and information space

Two main procedures: (1) Sampling nodes to expand (2) Communication & Extending

the nodes

2. Categorize existing nodes (of the same tree)

based on predefined criteria

(e.g. same robot-state)

1. Each node contains information about

the robot-state, uncertainty, communication-vector

3. Sample a group to expand its nodes

4. Sample a control input from the set of

admissible control inputs

13 of 18

Distributed Active Information Gathering

13

Basic Idea: Each robot builds its own local tree that explores both physical and information space

Two main procedures: (1) Sampling nodes to expand (2) Communication & Extending

the nodes

Apply sampled control input on robot dynamics and

compute new robot states for each node

14 of 18

Distributed AIG – Belief Propagation

14

  • Each robot updates its distribution (a-posteriori distribution) as a geometric mean of its

neighbors' beliefs and its own observation likelihood function

Sample one of the available neighbor’s nodes

The selected sampling functions affect the performance of the algorithm.

We propose biased sampling strategies that bias exploration towards informative areas.

15 of 18

Distributed AIG - Optimality Guarantees

15

Asymptotic Optimality

Probabilistic Completeness

16 of 18

Simulation Results

16

Fully Connected Communication Graph

Sparsely Connected Communication Graph

17 of 18

Scalability Analysis

17

robots

targets

18 of 18

18

Thank you!