Real-World Planning with PDDL+ and Beyond

Paper · Source
Task Planning

Real-world applications of AI Planning often require a highly expressive modeling language to accurately capture important intricacies of target systems. Hybrid systems are ubiquitous in the real-world, and PDDL+ is the standardized modeling language for capturing such systems as planning domains. PDDL+ enables accurate encoding of mixed discretecontinuous system dynamics, exogenous activity, and many other interesting features exhibited in realistic scenarios. However, the uptake in usage of PDDL+ has been slow and apprehensive, largely due to a general shortage of PDDL+ planning software, and rigid limitations of the few existing planners. To overcome this chasm, we present Nyx, a novel PDDL+ planner built to emphasize lightness, simplicity, and, most importantly, adaptability. The planner is designed to be effortlessly customizable to expand its capabilities well beyond the scope of PDDL+. As a result, Nyx can be tailored to virtually any potential real-world application requiring some form of AI Planning, paving the way for wider adoption of planning methods for solving real-world problems.

Introduction. Realistic planning problems require an expressive modeling language to accurately capture the innate intricacies of the modeled scenario. Indeed, most existing domainindependent planning languages, such a STRIPS (Fikes and Nilsson 1971), ADL (Pednault 1989), PDDL (McDermott et al. 1998), and even PDDL2.1 (Fox and Long 2003), lack features to describe commonplace elements of real-world systems. As a result, the vast majority of planning models are severely abstracted or limited in scope, often being forced to ignore aspects of the domain that, in the real world, have significant impact on the system’s operations such as environmental phenomena. PDDL+ (Fox and Long 2006) is one of the most expressive modeling languages and was designed to model mixed discrete/continuous (hybrid) systems. PDDL+ is, arguably, the closest AI Planning has come to accurately representing realistic scenarios as planning domains. It has proved useful for capturing a wide range challenging AI domains from traffic control(Vallati et al. 2016) and physicsbased games(Piotrowski et al.

Discussion / Conclusion. Real-world systems are challenging for AI with respect to modeling and solving them. Currently available PDDL+ planners have a steep learning curve and require expert knowledge. Nyx aims to increase the accessibility to AI planning, particularly for realistic feature-rich domains, by focusing on simplicity and adaptability. However, such advantages come at a cost of raw computational performance compared to complex optimized code and low-level languages such as C/C++. For transparency, Table 2 shows comparison with UPMurphi and ENHSP2 (both using blind BFS) in representative speed of state space exploration in nodes per second. Crucially, it should be noted that Nyx addresses a different classes of problems from the other planners. Our planner facilitates rapid prototyping of planner extensions and novel heuristics, solving emerging classes of domains which cannot be handled by any established planners, and reasoning with composite PDDL+ models with external functions & features.

Synthesis notes from this paper's topics 8

This paper has no direct semantic matches yet — here are notes from the topic(s) it's filed under.

Lines of inquiry this paper opens 24

Research framings built by reading the notes related to this paper — the questions it feeds into.

Can language models reliably simulate personas and predict behavior? Which reinforcement learning modifications most improve dialogue quality in language models? What makes process supervision effective for training complex reasoning models? What capabilities differentiate diffusion from autoregressive language models? Should models ask for clarification when facing ambiguous or under-specified information? What limits language model accuracy in evaluating ideas? What prevents LLMs from applying their reasoning knowledge to improve outputs? What explains the gap between benchmark scores and true reasoning capability? How do multi-agent systems fail when coordination breaks down? Does scaling reasoning capability create fundamental tradeoffs in control and reliability? How does fine-tuning trade off accuracy against reasoning quality? Can smaller specialized models match frontier models on key metrics? What gaps exist between benchmark performance and real deployment outcomes? When should retrieval systems decide to fetch new information? What human oversight must AI research systems have?