Click here to load reader

Multi-Sensor State Estimation on Dynamic Quadruped Robots...IROS 2016 Workshop on State Estimation and Terrain Perception Daejeon Convention Center - Daejeon, Korea Multi-Sensor State

  • View
    0

  • Download
    0

Embed Size (px)

Text of Multi-Sensor State Estimation on Dynamic Quadruped Robots...IROS 2016 Workshop on State Estimation...

  • IROS 2016 Workshop on State Estimation and Terrain PerceptionDaejeon Convention Center - Daejeon, Korea

    Multi-Sensor State Estimation on Dynamic Quadruped Robots

    Marco Camurri [email protected] Legged System Lab

    Istituto Italiano di Tecnologia – Genoa, Italy

  • October 14th, 2016 Marco Camurri 2/30

    Summary

    1) Hydraulic Quadruped (HyQ)– Description– Characteristic motions– Sensors

    2) State Estimation– Overview– Modules– Applications

    3) Mapping– Definition– Applications

  • October 14th, 2016 Marco Camurri 3/30

    Hydraulic Quadruped (HyQ)

  • October 14th, 2016 Marco Camurri 4/30

    Specifications

    ● 12 Degrees of Freedom● ~80 kg● 1 m x 0.5 m x ~0.8 m● Fully torque controlled● Fully hydraulic● 145 Nm (at 16 MPa)

    http://www.iit.it/hyq

  • October 14th, 2016 Marco Camurri 5/30

    Characteristic Motions

    ● Planned crawl● Trot● Flying trot● Chimney Climb

    https://www.youtube.com/HydraulicQuadruped

  • October 14th, 2016 Marco Camurri 6/30

    MiniHyQ, HyQ2Max, and...

    MiniHyQ HyQ2Max (source: Reuters)

    MOOG @ IIT Joint LabIntegrated Servo Actuators

    http://moog.iit.it

  • October 14th, 2016 Marco Camurri 7/30

    Sensors

    ● Microstrain GX3-25

    ● Optical encoders

    ● Load cells

    ● ASUS Xtion

    ● Multisense SL

    ● Hokuyo URG-04LX

  • October 14th, 2016 Marco Camurri 8/30

    State Estimation

  • October 14th, 2016 Marco Camurri 10/30

    Pronto

    ● Modular● EKF-based● History of meas.● Open Source*

    ● Used/tested for the DRC (MIT, ViGIR, …)

    ● LCM§ based § https://lcm-proj.github.io/Drift-free humanoid state estimation fusing kinematic, inertial and LIDAR sensingM. Fallon, M. Antone, N. Roy and S. Teller

    2014 IEEE-RAS International Conference on Humanoid Robots, Madrid, 2014 *https://github.com/ipab-slmc/pronto-distro

  • October 14th, 2016 Marco Camurri 11/30

    Modules

    ● Proprioceptive:– IMU (prediction)– Leg Odometry

    ● Exteroceptive:– Visual Odometry (FOVIS)– Gaussian Particle Filter (GPF)– Fast and Robust Scan Matcher (FRSM)– Vicon (ground truth)– ...

  • October 14th, 2016 Marco Camurri 12/30

    Modules

    ● Proprioceptive:– IMU (prediction) → bias, drift– Leg Odometry → drift, slippage, leg compliance

    ● Exteroceptive:– Visual Odometry (FOVIS) → featureless areas– Gaussian Particle Filter (GPF) → pre-acquired map– Fast and Robust Scan Matcher (FRSM) → only planar– Vicon (ground truth)– ...

  • October 14th, 2016 Marco Camurri 13/30

    State Estimation Scheme

    Stance Detection

    Leg Odometry

    IMU EKFω II , ẍ II

    ~̇x BBq , q̇ , τ

    ωBB , ẍBB , θB-W , ẋB-B , x B-W

    Controller

    ~s ∈B4

    Ethernet ~̇x BB , ~x BW ,~θBW

    ωBB , ẍBB , θB+W , ẋB+B , x B+W

    Clouds, LiDAR scans, ...GPFFRSM VO ...

    1 kHz

    500 Hz

    10-40 Hz

    250 kHz

    250 kHz

    250 kHz

    500 Hz

    250 kHz

    10-40 Hz

    500 Hz

  • October 14th, 2016 Marco Camurri 15/30

    Leg Odometry

    ● Ground Reaction Forces estimation● Stance Detection● Velocity computation● Covariance estimation

  • October 14th, 2016 Marco Camurri 16/30

    FRSM and Trunk Control

    ● RCF with push recovery

    ● Robot controlled to stay on target position

    ● Hokuyo URG 04-LX

    A reactive controller framework for quadrupedal locomotion on challenging terrainVictor Barasuol, Jonas Buchli, Claudio Semini, Marco Frigerio, Edson R De Pieri, Darwin G Caldwell2013 IEEE International Conference on Robotics and Automation (ICRA)

    RANGE - Robust Autonomous Navigation in GPS-denied EnvironmentsAbraham Bachrach, Samuel Prentice, Ruijie He Nicholas RoyJournal of Field Robotics, 2011

  • October 14th, 2016 Marco Camurri 17/30

    Gaussian Particle Filter

    ● Tested on Atlas/Drones

    ● Suitable for aggressive motions

    ● High Quality map required

    State estimation for aggressive flight in GPS-denied environments using onboard sensingA. Bry, A. Bachrach and N. Roy2012 IEEE International Conference on Robotics and Automation (ICRA), Saint Paul, MN, 2012

  • October 14th, 2016 Marco Camurri 18/30

    Gaussian Particle Filter

    IMU+Leg Odometry

    IMU+Leg Odometry+Gaussian Particle Filter

    Ground truthEstimate

    Ground truthEstimate

  • October 14th, 2016 Marco Camurri 19/30

    FOVIS

    ● Tested on Atlas/Drones

    ● Lightweight● Position or

    velocity measure

    Visual Odometry and Mapping for Autonomous Flight Using an RGB-D Camera. Albert S. Huang, Abraham Bachrach, Peter Henry, Michael Krainin, Daniel Maturana, Dieter Fox, and Nicholas Roy. Int. Symposium on Robotics Research (ISRR), Flagstaff, Arizona, USA, Aug. 2011

  • October 14th, 2016 Marco Camurri 20/30

    FOVIS

    ● Tested on Atlas/Drones

    ● Lightweight● Position or

    velocity measure

    Visual Odometry and Mapping for Autonomous Flight Using an RGB-D Camera. Albert S. Huang, Abraham Bachrach, Peter Henry, Michael Krainin, Daniel Maturana, Dieter Fox, and Nicholas Roy. Int. Symposium on Robotics Research (ISRR), Flagstaff, Arizona, USA, Aug. 2011

  • October 14th, 2016 Marco Camurri 21/30

    Selective ICP

    ● Selective ICP: register only the points in motion, geometrically relevant

    ● Fuse with IMU

    Real-time depth and inertial fusion for local slam on dynamic legged robots.M. Camurri, S. Bazeille, C. Semini, and D. G. CaldwellIEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), 2015

  • October 14th, 2016 Marco Camurri 22/30

    Selective ICP

    ● Frame-to-frame background subtraction

    ● Morphologic dilation

    ● Point cloud selection

    ● Iterative Closest Point (ICP) registration

    ● Black image (no edges) no →motion

    t-1 t

    Real-time depth and inertial fusion for local slam on dynamic legged robots.M. Camurri, S. Bazeille, C. Semini, and D. G. CaldwellIEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), 2015

  • October 14th, 2016 Marco Camurri 23/30

    Mapping

  • October 14th, 2016 Marco Camurri 24/30

    Robocentric Mapping

    ● Current sensed cloud is the most trustworthy● Current map should accumulate drift backwards

    ● Current map is the newest cloud plus previous map aligned to the current cloud

    ● Less accurate data is the oldest, and automatically discarded when out of scope

    M n=M(n−1)+ T n(n−1 ) ⋅Cn

    Real-time depth and inertial fusion for local slam on dynamic legged robots.M. Camurri, S. Bazeille, C. Semini, and D. G. CaldwellIEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI), 2015

  • October 14th, 2016 Marco Camurri 25/30

    Visual Pattern Classification

    ● Local heightmap around target footholds● Each heightmap is classified to select an offset correction

    on the touch down coordinateReactive trotting with foot placement corrections through visual pattern classificationV. Barasuol, M. Camurri, S. Bazeille, D. G. Caldwell and C. SeminiIntelligent Robots and Systems (IROS), 2015 IEEE/RSJ International Conference on, Hamburg, 2015

  • October 14th, 2016 Marco Camurri 27/30

    Octomap and planning

    On-line and On-board Planning and Perception for Quadrupedal LocomotionC. Mastalli, I. Havoutis, A. W. Winkler, D. G. Caldwell and C. SeminiIEEE International Conference on Technologies for Practical Robot Applications (TEPRA) 2015

    ● Scan with PTU● Scan Merging with

    Octomap● Feature extraction● Reward computation● Planning from reward

    map

  • October 14th, 2016 Marco Camurri 28/30

    Conclusions

    ● State Estimation is crucial for robot control, mapping and planning

    ● Multiple sources help being robust against more scenarios

    ● Local mapping helps keeping uncertainty away from where you want to operate

  • October 14th, 2016 Marco Camurri 29/30

    Alex Posatskiy Jose Colmenares

    Bilal Ur Rehman Carlos Mastalli

    Marco Camurri Yifu Gao

    Roy Featherstone

    Janne Koivumaki Elco Heijmink

    Claudio Semini Marco Frigerio Victor Barasuol Michele Focchi Andreea RadulescuRomeo Orsolino

    Sep Driessen

    The Dynamic Legged System Lab and friends:

  • October 14th, 2016 Marco Camurri 30/30

    Thanks for your attention!Questions?

    [email protected]

    Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 10Slide 11Slide 12Slide 13Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Slide 22Slide 23Slide 24Slide 25Slide 27Slide 28Slide 29Slide 30