Team Description: Pumas@Home 2027
Affiliation: Bio-Robotics Laboratory, Faculty of Engineering, National Autonomous University of Mexico (UNAM)
Participation: RoboCup@Home 2027 (Nuremberg, Germany) — Open Platform League (OPL) & Domestic Standard Platform League (DSPL)
Team Members: Jesús Savage, Rafael Cuéllar, Joshua Martínez, Marco Soriano, Sergio Cano, Angel Hernández, Jorge Méndez, Fidel Martínez, Elohim Ramírez, Martín Jiménez, and Marco Negrete.
1. About the Team & Track Record
Founded in 1996, the UNAM Bio-Robotics Laboratory focuses on developing service robots to assist people in home, office, and healthcare environments. Pumas@Home has regularly participated in RoboCup since Bremen 2006:
- Justina: Custom-built general-purpose domestic service robot developed in-house at UNAM.
- Takeshi (Toyota HSR): Domestic Standard Platform used to evaluate software modularity and adaptability on commercial systems.
- Awards & Achievements: Finalists in João Pessoa 2014, Hefei 2015, Montreal 2018, and Sydney 2019; best in Speech Recognition & NLU in Nagoya 2017; and awarded Best @Home Navigation consecutively since 2022.
2. Software Architecture: ViRBot System
Our software stack is built upon the four-layered ViRBot (Virtual Real Robot) architecture, implemented in ROS 2 Jazzy Jalisco:
- Inputs: Sensor signal processing and pattern recognition to model objects and people into symbolic environment beliefs.
- Knowledge Management: 2D SLAM occupancy maps, Kalman filter localization, and rule-based expert logic using NASA's CLIPS.
- Planning: Validates beliefs against knowledge models and activates sequences of actions to achieve assigned goals.
- Execution: Hardwired state machines executing skills (search, navigation, manipulation) under a continuous supervisor module capable of autonomous failure recovery.
3. Key Research & Novel Contributions (2027)
A. Hierarchical Supervised Execution Architecture
A three-level supervision framework written in C++ (ROS 2) ensuring system resilience in unstructured domestic environments:
- Global Supervisor: Tracks overall mission state and manages plan-level recovery strategies.
- General Executor: Coordinates concurrent execution across specialized modules (Vision, Navigation, Manipulation, HRI) while managing resource constraints and conflicts.
- Module Managers: Implement localized, autonomous failure recovery (e.g., retrying perception angles or alternate grasping strategies) before escalating errors.
B. Neuro-Symbolic Action Planning (Qwen-CLIPS)
A hybrid planner combining large language model semantic interpretation with rule-based predictability for the General Purpose Service Robot (GPSR) task:
- Fine-tuned Qwen2.5-7B via QLoRA over 40,000 GPSR command-goal pairs.
- Outputs structured JSON canonical goals (e.g., go, find, take, deliver).
- The goal parser transforms outputs into formal
gpsr-goalfacts in CLIPS, expanding them into validated, deterministic action plans.
C. Supervised Learning-Based Manipulation
Hybrid manipulation pipeline combining classical kinematic planning with Imitation Learning (IL), Reinforcement Learning (RL), and Vision-Language-Action (VLA) models:
- Rigid Articulated Objects: Autonomous door opening via model-based grasp planning and learned interaction policies.
- Deformable Materials: Autonomous shirt folding using teleoperated imitation learning demonstrations to bypass cloth simulation constraints.
D. Centralized Knowledge Management (Blackboard Architecture)
A centralized single-source-of-truth module using an SQLite relational database wrapped in ROS 2 services. Ensures persistent memory across boots (usual object locations, person embeddings) and enforces consistency constraints across simultaneous subsystems.
E. Pumas Navigation & Hybrid Neuroevolution
- Pumas Navigation: 2D SLAM maps, AMCL localization fusing wheel odometry and Hokuyo LiDAR, A* path planning, and Artificial Potential Fields (APF) for real-time human avoidance.
- Semantic Neuroevolution: Integrating lightweight CNN semantic obstacle classification with NEAT (NeuroEvolution of Augmenting Topologies) evolving LSTM controllers for robust mapless navigation in dynamic environments.
4. Robot Justina: Technical Specifications
| Mobile Base | Festo Robotino 4 (Omnidirectional, 1 m/s max speed) |
| Torso & Head | Electric linear actuator torso (Height: up to 1.65 m); 2-DOF expressive head |
| Manipulators | Dual 6-DOF anthropomorphic arms (5 kg maximum payload) |
| Sensors & Audio | Hokuyo URG-04LX-UG01 LiDAR, ASUS Xtion sensor, Røde Shotgun microphone, Bose speaker |
| Software Stack | ROS 2 Jazzy Jalisco, YOLO11, OWLv2, Florence-2, Whisper (ASR), Piper (TTS), MoveIt2, Pumas Navigation |
| Onboard Compute | Alienware m15 R6 (Intel Core 7 240H, 16 GiB RAM, 1.0 TB SSD, Ubuntu 24.04.5 LTS) |