Adaptive Planning: Dynamic Decision Making

• Adaptive Planning (AP) is a family of dynamic decision-making methods that revise strategies as tasks, environments, or assumptions change.
• It leverages structured state representations and unified planning languages like PDDL to efficiently reuse domain knowledge across varied scenarios.
• AP integrates heuristic search, model learning, and LLM-based model construction to enable responsive planning under uncertainty and evolving objectives.

Searching arXiv for recent and relevant papers on adaptive planning across AI planning, RL, robotics, and LLM-based planning. Adaptive Planning (AP) denotes a family of planning approaches in which behavior is revised, specialized, or generalized as tasks, goals, environments, or assumptions change. In AI and reinforcement learning, it is not a single formalism but a theme: planning systems that can exploit structured models, respond to uncertainty and time, and generalise or update behaviour across changing tasks and environments (Chen et al., 2024). In climate and institutional planning, AP is characterized as dynamically robust rather than optimally tailored to one future, emphasizing short-term actions that perform well now, keeping long-term options open, and pre-specifying contingency actions and monitoring systems to detect triggers or tipping points (Ross et al., 2022). Across these usages, AP is best understood as an orientation toward decision-making under model incompleteness, environmental change, and evolving objectives.

1. Definitions and formal scope

A standard AI planning baseline places AP inside the Markov Decision Process formalism

$M=\langle S,A,T,R,G\rangle,$

with terminal goal states $G$, a policy $\pi:S\to A$, and the usual objective of maximizing value; what distinguishes AP is not a different objective, but structured access to $T$ and $R$ and the ability to exploit that structure efficiently (Chen et al., 2024). A complementary classical-planning formulation writes a planning model as

$$\mathcal{M}=\langle \mathcal{S}, s^I, s^G, \mathcal{A}, \mathcal{T}\rangle,$$

where plans are action sequences that transform the initial state into a goal state under symbolic transitions (Tantakoun et al., 22 Mar 2025).

A central distinction in the planning survey is a hierarchy of access to the environment: continuing, episodic, generative, analytic, and structured. AP is placed at the structured level, where analytic access is augmented with a compact, symbolic, factored representation of transitions and rewards (Chen et al., 2024). This makes AP especially relevant when the environment is not treated as a black box, but as a domain whose compositional structure can be reused across many tasks.

Outside core AI, adaptive planning is defined against traditional static planning. Traditional hazard planning is described as producing a static optimal plan for a single “most likely” future or a small set of scenarios, whereas adaptive planning is built for deep uncertainty, with a shared strategic vision, short-term actions, long-term option preservation, and contingency actions tied to monitoring systems (Ross et al., 2022). This suggests that the common denominator of AP is not a single algorithmic template but the explicit management of change, uncertainty, and model revision.

2. Structured representations and planning languages

Classical AP relies on symbolic, factored state representations. States are sets of grounded propositions under the closed-world assumption; predicates and objects generate propositions, goals are propositional formulas, and action schemata have parameters, preconditions, add effects, and delete effects (Chen et al., 2024). For a grounded action $a$ applicable in state $s$, the successor is

$$s'=(s\setminus \mathrm{del}(a))\cup \mathrm{add}(a),$$

which compactly defines an exponentially large state graph. This factorization is one of the main structural sources of adaptivity, because planners can focus on relevant facts and generalize over varying object sets.

The dominant representation language is PDDL, which separates a domain file from a problem file. The domain file contains type hierarchy, predicates, and action schemas; the problem file contains objects, initial state, and goal formula (Tantakoun et al., 22 Mar 2025). The survey on AI planning extends this core with PPDDL for probabilistic effects, RDDL for factored MDPs and exogenous processes, PDDL 2.1 for numerics and temporality, and PDDL+ for continuous variables, processes, and events (Chen et al., 2024). These languages widen AP from deterministic symbolic search to uncertainty-aware, time-aware, and hybrid dynamical planning.

The same literature also places hierarchical task networks, temporal logic control knowledge, reward machines, and multi-agent encodings such as MA-STRIPS within the broader AP toolbox (Chen et al., 2024). In each case, adaptation comes from explicit structure: decomposition, temporally extended constraints, or coordination mechanisms that can be reused across related tasks rather than re-engineered instance by instance.

3. Adaptive mechanisms during planning and execution

At the algorithmic level, AP is strongly associated with heuristic search over structured state spaces. Heuristics derived from delete relaxation, abstraction, cost partitioning, landmarks, novelty, and width-based search make it possible to replan quickly when goals or initial conditions change (Chen et al., 2024). A*, Greedy Best-First Search, Weighted A*, anytime search, LAO*, LRTDP, and UCT provide different trade-offs among optimality, speed, and online responsiveness; the important point for AP is that these methods exploit reusable domain-level structure rather than solving each instance from scratch.

Several works make the adaptive mechanism explicit. “Adaptive Online Planning” combines MPC-style planning with model-free learning and uses uncertainty estimates from a value ensemble together with predicted planning benefit to decide how far and how much to plan in an online, reset-free, non-stationary setting (Lu et al., 2019). “Adaptive Planning with Generative Models under Uncertainty” uses a threshold-based replanning policy driven by predictive uncertainty from a Deep Ensemble of inverse dynamics models; in locomotion tasks, this reduces replanning frequency to only about 10% of the steps without compromising performance (Jutras-Dubé et al., 2024). “Path Planning in Dynamic Environments with Adaptive Dimensionality” adapts the search representation itself, considering the time dimension only in regions where a potential collision may occur and planning in a low-dimensional state-space elsewhere; the method is complete and guaranteed to find a solution, if one exists, within a cost sub-optimality bound (Vemula et al., 2016).

A different adaptive mechanism appears in “Multi-tier Automated Planning for Adaptive Behavior,” where ordered FOND domains encode different sets of assumptions and corresponding goals. A multi-tier planning problem

$M=\langle \langle \Omega,\le\rangle, s_I, G\rangle$

supports adaptation by switching to lower tiers when observations violate higher-tier assumptions; the compilation to dual FOND planning justifies the use of both fair and unfair actions (Ciolek et al., 2020). Here adaptation is model switching under assumption failure rather than heuristic replanning alone.

4. Learning, generalization, and language-model mediation

A major contemporary strand treats AP as learning reusable structure. The planning survey frames generalisation in planning through a domain shared across training and test tasks, with a learner producing a knowledge artifact and a planner using that artifact on unseen tasks (Chen et al., 2024). Model learning from state–action traces, images, or active interaction upgrades black-box environments to structured planning domains; learning from structured domains then yields heuristics, policies, preprocessing rules, or generalized plans that transfer across object counts and task instances (Chen et al., 2024).

The neuro-symbolic line extends this by using LLMs as planning modelers rather than end-to-end planners. The LLM survey argues that LLMs are more reliable as modelers: they can extract predicates, infer action semantics, generate domain/problem files, repair models, and update task or domain specifications online as goals and observations change (Tantakoun et al., 22 Mar 2025). This role is central to AP because it turns changing language, documents, and feedback into revised symbolic models that classical planners can solve.

AdaPlanner supplies an explicitly feedback-driven LLM realization of AP. It introduces a closed-loop method in which an LLM agent refines its self-generated plan adaptively in response to environmental feedback using in-plan and out-of-plan refinement, employs a code-style prompt structure to mitigate hallucination, and uses skill discovery to reuse successful plans as few-shot exemplars (Sun et al., 2023). In ALFWorld and MiniWoB++, it outperforms state-of-the-art baselines by 3.73% and 4.11% while utilizing 2x and 600x fewer samples, respectively (Sun et al., 2023). The broader implication is that explicit plan revision, rather than stepwise reaction alone, is becoming a central mechanism for adaptive behavior in LLM agents.

5. Representative application domains

Robotics provides several concrete realizations of AP. “EVA-Planner” adapts quadrotor flight aggressiveness online according to obstacle distribution and quadrotor state by using an environmental adaptive safety aware method inside a multi-layered MPCC framework, so the vehicle flies slower in denser environments and increases its speed in a safer area (Quan et al., 2020). “APPLR” and the broader “APPL” framework adapt classical navigation stacks by learning parameter policies over quantities such as maximum speed, sampling rates, obstacle-cost weights, and inflation radius, using reinforcement learning, demonstrations, interventions, evaluative feedback, or combinations thereof (Xu et al., 2020, Xiao et al., 2021). “Adaptive Dynamics Planning” adapts the fidelity of the robot’s dynamics model online during planning, using reinforcement learning to choose how detailed the dynamics modeling should be based on current environment and navigation status (Yuanjie et al., 6 Oct 2025). “Socially Adaptive Path Planning Based on Generative Adversarial Network” embeds a learned socially aware cost model into RRT*, improving the anthropomorphic degree of robot motion planning and the homotopy rate between planned and demonstration paths (Wang et al., 2024).

Mission-level routing under uncertainty supplies another canonical AP setting. “Adaptive Probabilistic Planning for the Uncertain and Dynamic Orienteering Problem” introduces UDOP and ADAPT, which iteratively performs execution and online planning based on an initial offline solution, updates uncertain travel-cost models with Bayesian inference, and chooses paths using safety beliefs; in UAV charging scheduling, it achieves a 100% Mission Success Rate across all tested scenarios (Qian et al., 2024). In network planning, “OmniPlan” interprets natural-language intents into a user-preference vector, dynamically selects among MIP, heuristic, and DRL experts, and reconfigures optimization weights online, achieving near-optimal and low-execution-time offloading while reducing latency by up to 97.8% and network device resource consumption by up to 11.5% (Zhu et al., 16 Jun 2026).

AP also appears outside engineering control. In flood planning for the Lower Rio Grande Valley, adaptive planning is defined through uncertainty acceptance, diverse participation, and development of co-benefits, yet the empirical analysis finds that hazard plans and discussions are largely lacking an adaptive approach (Ross et al., 2022). This non-AI literature shows that AP depends not only on algorithms but also on monitoring regimes, participation mechanisms, funding rules, and institutional willingness to maintain contingent pathways.

6. Limitations, misconceptions, and open directions

A persistent misconception is that AP is merely continuous replanning. The surveyed literature points to a broader picture: AP may involve structured abstractions, uncertainty-triggered horizon control, model acquisition, generalized planning, parameter adaptation, tier switching, or planner selection, all under changing tasks or assumptions (Chen et al., 2024). A second misconception is that AP and RL are interchangeable. The planning survey explicitly positions AP as starting from structured models or learning them, while RL contributes machinery for model, heuristic, or policy acquisition when models are not given (Chen et al., 2024).

Current systems retain strong limitations. LLM-based model construction still suffers from hallucinations, semantic errors, and weak support for complex temporal, numeric, and probabilistic constraints at scale (Tantakoun et al., 22 Mar 2025). Generative-model AP depends on the calibration of inverse dynamics uncertainty and the quality of the diffusion model (Jutras-Dubé et al., 2024). Adaptive probabilistic routing assumes normality and a Normal–Gamma prior for power consumption (Qian et al., 2024). Adaptive dimensionality planning assumes known dynamic-obstacle trajectories (Vemula et al., 2016). Multi-tier compilation is computationally demanding and motivates stronger planning systems for dual fairness assumptions (Ciolek et al., 2020).

The dominant research direction is integrative. The planning and LLM surveys both point toward systems that combine model acquisition, structural analysis, heuristic search, learned guidance, and richer symbolic languages extending beyond classical PDDL to temporal, numeric, conditional, and stochastic models (Chen et al., 2024, Tantakoun et al., 22 Mar 2025). A broader implication is that AP research is converging on two coupled questions: how to maintain adaptable decision procedures under uncertainty, and how to maintain the models, monitoring, and organizational conditions that let those procedures switch goals, pathways, or fidelities when reality departs from assumptions.