Team Description: Pumas@Home 2024
Affiliation: Bio-Robotics Laboratory, Faculty of Engineering, National Autonomous University of Mexico (UNAM)
Participation: RoboCup@Home 2024 — Domestic Standard Platform League (DSPL)
Team Members: Oscar Fuentes, Jesús Savage, Marco Negrete, Rubén Sandoval, Oscar Durón, Joel Gago, and Daniel Vanegas.
1. About the Team & Track Record
Pumas Team from UNAM conducts research combining symbolic AI with digital signal processing techniques to develop reliable domestic service robots. The team focuses on validating and porting its software framework to commercial platforms:
- Platform: Takeshi (Toyota Human Support Robot - HSR).
- Past Accomplishments: 2nd place in RoboCup@Home DSPL (Montreal 2018) and 4th place (Sydney 2019), followed by participation in the World Robot Summit and Rockin competitions.
2. Software Architecture: ViRBot System
The robot operates under the ViRBot (Virtual and Real Robot system) layered architecture, coordinating low-level processing with high-level cognitive decision making:
- Inputs Layer: Reads raw data from the Hokuyo laser scanner, RGB-D sensors, stereo cameras, wide-angle cameras, and microphone arrays. Employs digital signal processing to construct symbolic beliefs about the environment.
- Knowledge Management Layer: Maintains environmental representations through SLAM techniques, Kalman Filter localization, and a rule-based expert system implemented in NASA's CLIPS. Links Linux ROS nodes with Windows modules via a custom Blackboard middleware.
- Planning Layer: Validates perceptual beliefs against stored knowledge, identifies situational contexts, and formulates goal sequences.
- Execution Layer: Deploys finite state machines for modular procedures (person finding, grasping, object transport) driven by action plans.
3. Current Research & Technical Challenges
A. The 4S Philosophy (Speed, Smooth/Smart, Stable, Safe)
- Speed: Accelerated perception and neural processing via onboard GPU (Nvidia GT 1080 / Quadro P4000) running YOLO, FaceNet, and OpenPose in real time.
- Smooth / Smart: Seamless Human-Robot Interaction (HRI) supporting multimodal spoken voice and visual inputs.
- Stable: Decoupled, modular architecture where independent ROS packages allow subsystem modification without system degradation.
- Safe: Redundant obstacle detection using RGB-D point clouds, physical bumper layers, and local reactive collision avoidance.
B. Integrated Navigation Framework
Navigation in domestic spaces combines global path planning with reactive layers:
- Global path routing via the A* algorithm on 2D occupancy grids through ROS AMCL.
- Point-cloud processing to isolate non-planar 3D obstacles (such as tables and chairs) and trigger route replanning.
- Local avoidance utilizing Artificial Potential Fields (APF) for real-time evasive maneuvers around moving individuals without complete trajectory resets.
- Localization recovery routines utilizing Hidden Markov Models (HMM) and laser scans.
C. Synthetic Training for Object Detection & Autonomous Manipulation
A training pipeline designed to bridge the reality gap without manual dataset labeling:
- Automated video frame segmentation and synthetic scene synthesis generating over 20,000 annotated images.
- Integration with YOLOv3 and statistical filtering for 3D point cloud orientation via Principal Component Analysis (PCA) to align the robotic gripper during grasping.
4. Open-Source Software Stack
| Package | Application / Role |
|---|---|
| ROS AMCL & Hector Map | 2D SLAM and map-based probabilistic localization |
| YOLO (Darknet) | Real-time 2D object detection trained with synthetic datasets |
| FaceNet & OpenPose | Facial identification and human pose detection |
| OpenCV | Image processing, filtering, and PCA rotation estimation |
| Vosk | Spoken command and speech recognition |
| NASA CLIPS | Rule-based expert reasoning and state-space action planning |