Team Description: Pumas@Home 2019
Affiliation: Bio-Robotics Laboratory, School of Engineering, National Autonomous University of Mexico (UNAM)
Participation: RoboCup@Home 2019 & ROCKIn — Open Platform League (OPL)
Team Members: Jesús Savage, Reynaldo Martell, Hugo Estrada, Julio Cruz, Marco Negrete, Jaime Márquez, Manuel Pano, Julio Martínez, Jesús Cruz, and José Cruz.
1. About the Team & Track Record
The Bio-Robotics Laboratory at UNAM has competed continuously in RoboCup@Home since Bremen 2006. Achievements include 2nd place in RoboCup 2018 (Montreal), 4th place and the Best in Speech Recognition and Natural Language Understanding award in Nagoya 2017, as well as regular participation in the European ROCKIn challenge.
2. Software Architecture: ViRBot System
Justina's software operates under the layered ViRBot (Virtual and Real Robot) architecture, coordinating ROS nodes on Linux with Windows modules through a custom Blackboard middleware:
- Inputs Layer: Acquires raw sensory data from Hokuyo LiDAR, sonar, infrared sensors, Microsoft Kinect RGB-D, and microphones. Signal processing and pattern recognition algorithms convert these streams into symbolic beliefs about the environment.
- Knowledge Management Layer: Manages SLAM-built maps, Kalman filter localization, and rule-based expert knowledge encoded with NASA's CLIPS.
- Planning Layer: Validates perceptual hypotheses against the knowledge base, recognizes situations, and determines physical operation sequences.
- Execution Layer: Deploys hardwired finite state machines to perform modular tasks such as person tracking, navigation, and manipulation.
3. Key Research & Novel Contributions
A. Natural Language Understanding via Conceptual Dependency (CD)
Implementation of Schank's Conceptual Dependency (CD) theory within an expert system to extract structured meaning from unstructured speech commands:
- Decomposes sentences into action primitives, notably ATRANS (transfer of abstract relationships, e.g., give, take) and PTRANS (transfer of physical location, e.g., go).
- Fills conceptual roles (Actor, Act, Object, Direction, and State) without requiring rigid grammatical rules.
- Feeds facts directly into the CLIPS inference engine to generate deterministic state-space action plans.
B. Intelligent Flat-and-Textureless Object Manipulation
A specialized vision and grasping pipeline developed for the Procter & Gamble Dishwasher Challenge (PGDC), dealing with highly reflective, thin, and textureless tableware:
- Applies HSV color-space thresholding followed by morphological closing and dilation to isolate reflective surfaces.
- Employs convex hull operations to produce clean 2D bounding boxes and masks under challenging lighting conditions.
- Combines 2D image areas with 3D point cloud metrics (point density, eigenvalues, and eigenvectors) to categorize objects into four classes: glass, dish, bowl, and cutlery.
- Calculates centroids and 3D roll, pitch, and yaw grasping angles, incorporating visual feedback to verify successful acquisition.
4. Robot Justina: Technical Specifications
| Mobile Base | Omnidirectional base driven by differential pair configuration and omnidirectional wheels |
| Torso & Head | 1-DOF elevating torso (worm screw and gear arrangement); 2-DOF pan-and-tilt head powered by Dynamixel servomotors |
| Manipulators | Two 7-DOF anthropomorphic arms actuated by 10 Dynamixel servomotors each |
| Sensors & Cameras | Hokuyo URG-04LX-UG01 LiDAR, Microsoft Kinect RGB-D sensor, Logitech Pro C920 Full HD camera |
| Audio Hardware | Røde NTG2 directional microphone, 4-microphone sound source localization array, dual speakers |
| Software Stack | ROS Kinetic & Blackboard middleware, ROS Gmapping, A* navigation, YOLO, FaceNet, OpenPoses, Loquendo TTS, NASA CLIPS |
| Embedded Compute | NVIDIA Jetson TX2 Developer Kit for GPU image acceleration, Ubuntu 16.04 LTS and Windows 7 VM |