Human-centered manipulation and navigation with Robot DE NIRO Towards Robots that Exhibit Manipulation Intelligence IROS workshop October 1, 2018 Nico Smuts 1 Sagar Doshi 1 Fabian Falck 1 Petar Kormushev 1 Kim Rants 1 John Lingi 1 1 Department of Computing, Imperial College London 2 Dyson School of Design Engineering, Imperial College London Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 1
Robot DE NIRO – Motivation and Hardware Design Microsoft Kinect depth camera • Macrosocial trends of aging and Stereovision cameras long-lived populations • Related work: focused on helping the elderly live independently [1] Baxter robot arms [2] [3] [4] • Ethical concerns [5] [6]: – human isolation – loss of control and personal Mount position of 2D liberty Hokuyo LIDAR – deception and infantilization • DE NIRO: “Support the QUICKIE base with feedback controller supporter” (the caregiver) and offer direct human-robot interaction with the care recipient Sources: [1] Care-O-Bot 3 [2] ASIMO [3] HRP-3 [4] DOMEO RobuMate [5] Sparrow et al. [6] Wallach et al. Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 2
Software implementation and Play Fetch routine • ROS as middleware • Finite-state machine to manage the control flow Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 3
Perception and user interaction • Face recognition : ResNet model pre-trained on faces applied to video frames retrieved by the Kinect camera, reaching an accuracy of 99.38% on a standard benchmark [1] • Speech recognition : Offline library CMU Sphinx [2]. Jspeech Grammar to allow reliable voice commands in a specific format. • Speech output : eSpeak [3] yielding a high reliability, rapid response time and an offline implementation. <object> <article> <command> fetch me a coffee give me an water the apple juice Sources: [1] “dlib face recognition documentation,” http://dlib.net/dnn [2] https://cmusphinx.github.io/wiki/ [3] http://espeak.sourceforge.net/index.html Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 4
Object Recognition and Manipulation • Target objects are localized using 2D fiducial markers [1] • Inverse kinematics solver to compute each of the seven joint angle trajectories needed to reach an object [2] • Safety: dynamic awareness procedure reacts to changes of the location of the target object during grasping and actively avoids collisions Sources: [1] https://github.com/chili-epfl/ros_markers [2] http://sdk.rethinkrobotics.com/wiki/IK_Service_-_Code_Walkthrough Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 5
Navigation stack • Static mapping : SLAM-based approach using the LIDAR sensor to detect spatial boundaries and 2D artifacts [1] • Localization : Dynamic map overlaid onto static map (for collision avoidance) • Trajectory planning : “Timed elastic band” approach [2] [3]. Sources: [1] http://wiki.ros.org/hector_mapping [2][3] http://wiki.ros.org/teb_local_planner Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 6
Conclusion Limitations Future work • nonholonomic design • increased awareness and safety through 360-degree camera rig • maximum payload of 2.2 kg per arm • 3D LIDAR • currently limited to forward motion only due to limited sensor capabilities • teleoperation through virtual reality (possible deadlocks in corners) headset and body tracking markers Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 7
Thank you! Code + Documentation + Videos https://www.imperial.ac.uk/robot-intelligence/software Fabian Falck – Robot DE NIRO – IROS workshop on Manipulation Intelligence 01/10/2018 8
Recommend
More recommend
