Introduction
Industry 4.0 is driving a transition from rigid, hard-coded factory floors toward flexible, "agent-aware" collaborative environments. In these smart factories, humans and robots (cobots) must work in tandem to handle mass customization. However, the challenge lies in dynamically determining the optimal agent for a given task based on physical capabilities, safety, and available information. This paper presents a framework that integrates Planning Domain Definition Language (PDDL) with a Partial Order Planning Forwards (POPF) temporal planner and a novel cost-based allocation algorithm to solve this problem.
System Architecture and Methodology
The proposed system functions as a high-level "commander" that orchestrates task sequences. It relies on three primary components to create a logical and physical bridge between planning and execution:
- PDDL (Planning Domain Definition Language): A standardized AI language used to define the "world" (the domain) and the specific "problem" (initial state vs. goal state).
- POPF (Partial Order Planning Forwards): A temporal planner that doesn't just look for a valid sequence of actions but searches for the cheapest one based on assigned costs.
- Task Allocation Algorithm: The core innovation where the authors define the "cost" of an action ($C_{Agent}$). This cost is calculated based on feasibility (e.g., physical strength), reachability (using capability maps to avoid awkward joint postures), and safety (penalizing actions that force robots near humans).
Information as a Resource
One of the most elegant ideas in this framework is treating "missing information" as an infinite cost. If a robot lacks the spatial coordinates for a part, the cost of a "PICK" action is set to infinity. This forces the planner to select a "COOPERATE" action where a human guides the robot. Once the robot records the necessary coordinates, the internal PDDL domain is updated, and the cost for autonomous action drops to zero for the next cycle, allowing the robot to take over autonomously.
Experimental Validation
To validate the framework, the researchers developed a physical testbed consisting of two 7-Degree-of-Freedom (DOF) Franka Emika robotic manipulators and one human worker. The task involved assembling "Basement parts" (135g) and "Rings" (73g) and moving them to a storage area. Human workers signaled task completion via a speech-to-text interface.
Key Results and Insights
The authors demonstrated the system's efficacy through two primary phases:
- The Learning Phase: Initially, a robot lacked assembly coordinates. The POPF solver generated a plan where the human worker guided the robot—a high-cost but necessary collaborative action.
- The Autonomous Phase: In the second production cycle, the system updated its domain with the acquired coordinates. The cost for the robot to work alone dropped significantly, and the planner automatically reassigned the task to the robot, freeing the human for other work.
Quantitative analysis showed that specific action costs (e.g., a "MOVE" action) varied significantly (from 0.637 to 10.09) depending on the agent's reachability and the target location, proving the system's sensitivity to physical constraints.
Limitations and Future Work
While successful, the authors note that the current cost function does not explicitly account for the time taken to perform an action or the precision (mean square error) of robot movements. Future work aims to include mobile robots and scouting drones, as well as integrating more complex path-planning algorithms to further reduce collision risks in crowded workspaces.
Reference: Chen, Q., & Pan, Y.-J. (2024). An Optimal Task Planning and Agent-aware Allocation Algorithm in Collaborative Tasks Combining with PDDL and POPF. arXiv:2407.08534v1.
Summary updated