Abstract
In this paper, we present an approach for the control of autonomous robots based on Automated Planning (AP) techniques, specifically through the development of the ROPEM (RObot Plan Execution with Monitoring) architecture. This architecture integrates deliberation with a standard planner, execution, monitoring, and replanning. By utilizing a low-level control layer implemented in the Player framework, we decouple high-level task execution from specific robotic hardware, ensuring platform independence and reusability. Given the non-deterministic nature of robot task execution, a dedicated monitoring module supervises the process to detect goal achievement or unexpected states. We evaluated ROPEM using the International Planning Competition (IPC) Rovers domain with a real Pioneer P3DX robot, testing its planning and replanning capabilities under various sources of uncertainty.
Introduction
While Automated Planning (AP) has seen significant theoretical advancement, its application to physical robotics remains challenging. Traditional robotic systems often rely on reactive control—pre-programmed responses to sensor inputs—which lacks the long-term reasoning necessary for complex, multi-stage missions.
Applying symbolic planning to the real world faces three primary hurdles: environmental uncertainty, where actions rarely produce exact intended results; modeling complexity, as capturing physical nuances in PDDL (Planning Domain Definition Language) is difficult; and platform dependency, where deliberation logic is often too tightly coupled with specific hardware. The ROPEM architecture addresses these by creating a robust bridge between abstract reasoning and physical execution.
The ROPEM Architecture
The ROPEM architecture is structured into four functional pillars designed to manage the flow from high-level goals to low-level actuator commands:
- Deliberation (Planning): Employs the SAYPHI planner to generate high-level action sequences from PDDL domain and problem definitions.
- Execution: Decomposes abstract plan steps (e.g., "Navigate to Waypoint") into discrete low-level behaviors such as
turnLeftormoveForward. - Monitoring: Acts as the system's "glue," continuously translating low-level sensor data back into high-level PDDL predicates. If a precondition for a planned action is violated, it triggers a replanning request.
- Low-Level Control: Uses the Player framework to abstract hardware specifics, allowing the deliberation logic to remain reusable across different robotic platforms.
This "Plan-Execute-Monitor" loop ensures that if the robot encounters an unexpected state or a failed action, it can immediately generate a new PDDL problem based on its current state and invoke the planner to find a new path to the goal.
Experimental Setup
The architecture was evaluated using the IPC Rovers domain, which simulates Mars exploration tasks like traversing waypoints, collecting samples, and transmitting data. Experiments were conducted on a Pioneer P3DX mobile robot equipped with sonar, bumpers, and a PTZ camera. The laboratory environment was mapped to a grid, and terrain variations (e.g., sandy vs. rocky) were simulated through software-induced delays to test the system's temporal reasoning.
Results and Discussion
The system demonstrated high efficiency and resilience under varying conditions. In simple navigation tasks, planning and replanning times were consistently low, typically under 0.5 seconds, with a general upper bound of less than 2 seconds across all tests.
When obstacles were introduced, the robot successfully completed its goals despite a significant increase in replanning episodes (averaging 20 per run). The monitoring system proved vital at higher speeds; although increased speed led to greater odometry noise and frequent waypoint overshooting, the monitor detected these errors and triggered corrective replanning. However, the researchers noted that cumulative odometry drift remains a bottleneck for scalability in larger environments.
Conclusion
ROPEM successfully demonstrates that decoupling high-level deliberation from low-level control, supported by a continuous monitoring loop, provides a robust framework for autonomous robot navigation in uncertain environments. Future work will focus on integrating more advanced localization techniques like SLAM to mitigate odometry drift and exploring Machine Learning to automatically refine PDDL models based on execution experience.
Summary updated