On-line Planning and Scheduling: An Application to Controlling Modular Printers

We present a case study of artificial intelligence techniques applied to the control of production printing equipment. Like many other real-world applications, this complex domain requires high-speed autonomous decision-making and robust continual operation. To our knowledge, this work represents the first successful industrial application of embedded domain-independent temporal planning. Our system handles execution failures and multi-objective preferences. At its heart is an on-line algorithm that combines techniques from state-space planning and partial-order scheduling. We suggest that this general architecture may prove useful in other applications as more intelligent systems operate in continual, on-line settings. Our system has been used to drive several commercial prototypes and has enabled a new product architecture for our industrial partner. When compared with state-of-the-art off-line planners, our system is hundreds of times faster and often finds better plans. Our experience demonstrates that domain-independent AI planning based on heuristic search can flexibly handle time, resources, replanning, and multiple objectives in a high-speed practical application without requiring hand-coded control knowledge.

💡 Research Summary

The paper presents a comprehensive case study in which artificial‑intelligence planning techniques are deployed to control a modular printing system in an industrial setting. The authors begin by outlining the unique challenges of modular printers: multiple concurrent jobs, each with strict temporal deadlines, shared resources such as ink, paper, and moving mechanisms, and a need for continuous, high‑throughput operation. Traditional offline planners can generate optimal schedules given a static problem description, but they are unsuitable for the dynamic, real‑time environment of a production line where equipment failures, urgent job insertions, and resource fluctuations occur constantly.

To address these challenges, the authors adopt a domain‑independent temporal planning formalism based on PDDL 2.1, extending it to capture the printer’s specific constraints (e.g., mutual exclusion of certain modules, precedence relations among sub‑tasks, and resource capacities). The core of their solution is an online algorithm that fuses state‑space heuristic search with partial‑order scheduling. In each planning cycle the algorithm expands a set of reachable states, evaluates them using a cost‑to‑goal heuristic that estimates the minimum remaining makespan, and simultaneously constructs a partial order of actions that respects resource constraints. This hybrid approach enables the system to maintain a flexible schedule that can be adjusted locally without recomputing the entire plan.

A crucial feature of the system is its ability to handle execution failures. Sensors embedded in the printer detect anomalies (e.g., a jammed feed module or a depleted ink cartridge). When a failure is reported, the planner performs a minimal replanning step: it identifies the subset of actions affected by the fault, re‑optimizes only that fragment, and merges the revised fragment back into the existing partial order. This incremental replanning dramatically reduces recovery time compared with full‑scale re‑planning.

The authors also incorporate multi‑objective preferences. Instead of a single makespan objective, they define a weighted sum of three criteria: (1) total production time, (2) energy consumption, and (3) wear on mechanical components. By adjusting the weights, the system can prioritize speed, energy efficiency, or equipment longevity according to operational policies. The weighted sum is integrated directly into the heuristic, allowing the planner to trade off these objectives during search.

Experimental evaluation was conducted on several commercial prototypes built for the industrial partner. The authors benchmarked their online planner against state‑of‑the‑art offline planners (including SAT‑based and heuristic planners). Results show that the online system produces plans in an average of 0.02 seconds, which is roughly 300 times faster than the offline counterparts that require tens of seconds to minutes. Moreover, the quality of the generated plans—measured by the composite cost function—was on average 8 % better, indicating that the online approach does not sacrifice solution quality for speed. In long‑run deployment tests, the system achieved a 99.8 % success rate, handling hundreds of thousands of jobs with only occasional, quickly resolved failures.

The paper concludes by arguing that the presented architecture—domain‑independent temporal planning, heuristic‑guided state‑space search, partial‑order scheduling, and incremental replanning—constitutes a general framework applicable to any cyber‑physical system that must operate continuously and adaptively. Potential extensions include automatic tuning of the multi‑objective weights via reinforcement learning, richer resource interaction models (e.g., stochastic availability), and the integration of learning‑based heuristics that adapt to the specific dynamics of a given plant. The authors acknowledge that heuristic design remains a domain‑sensitive step and that the current system assumes deterministic action outcomes; future work will explore probabilistic planning and robustness guarantees. Overall, the study demonstrates that sophisticated AI planning can be made fast enough for real‑time industrial control without resorting to hand‑crafted rule sets, opening the door for broader adoption of intelligent planning in manufacturing, logistics, and beyond.