Task Planning & Execution

• Task planning and execution is a process that converts abstract goals into structured, sequential actions using symbolic, hybrid, and LLM-based methods.
• It leverages hierarchical strategies like skill-centric decomposition and task-decoupled planning to optimize decisions and handle uncertainties in real-time.
• Empirical benchmarks show improvements in execution time and success rates, underscoring its practical significance in robotics and AI applications.

Task planning and execution encompasses the holistic process by which intelligent agents, particularly in robotics and LLM–driven systems, synthesize and translate abstract goals into concrete sequences of actions that attain those goals in real or simulated environments. This involves constructing, maintaining, and iteratively updating symbolic or skill-based action plans, selecting appropriate execution strategies, and robustly adapting to unmodeled contingencies or dynamic contexts during run time. Recent research has advanced from monolithic, stepwise planning paradigms to continually re-planned, hierarchical, and skill-centric architectures that emphasize efficiency, generalizability, and the capacity to handle ambiguous or partially grounded domains.

1. Formal Paradigms and Frameworks

Task planning and execution can be instantiated in several foundational paradigms, ranging from classical symbolic planning to skill-decomposition and LLM-driven agent frameworks:

• Symbolic Planning: Classical models define a tuple of states $S$, actions $A$, goal set $G$, transition functions (deterministic or stochastic), and cost functions $c(s, a)$. The objective is to compute a plan $\pi = [a_1, ..., a_n]$ that transitions an initial state $s_0$ to a state $s_g \in G$, often minimizing cumulative cost (Arora et al., 4 Feb 2025).
• Hybrid Task and Motion Planning (TAMP): TAMP frameworks separate high-level symbolic task planning from continuous motion generation, accommodating constraints where some action parameters cannot be grounded until execution time, and bridging gaps via closed-loop behaviors (Pan et al., 2024).
• Skill-Centric and Hierarchical Decomposition: Skill-centric systems break down tasks into meta-skills, which are themselves grounded by skill models or policies. Hierarchical approaches utilize a scheduling layer (often LLM-driven), a meta-skill layer (vision-language-action models or other neural policies), and low-level controllers, enabling dynamic manipulation of novel objects and tasks (Mao et al., 2024).
• LLM-Based Global Planning and Hierarchical Execution: Architectures such as GoalAct introduce a continuously updated global plan $G$, decomposed into high-level skills and detailed by a hierarchical executor at each step, providing robust adaptability in real-world domains (Chen et al., 23 Apr 2025).
• Task-Decoupled Planning (TDP): TDP employs a directed acyclic graph (DAG) of sub-goals managed by a Supervisor. Each sub-task is solved in isolation using a scoped context, confining reasoning, execution, and replanning to local sub-tasks to prevent error propagation (Li et al., 12 Jan 2026).

2. Planning and Execution Mechanisms

Global and Local Planning

• Global Plan Updating: Some systems recompute a complete high-level action plan at every step, e.g., $G_t = \pi(Q, T, S_t)$, where $Q$ is the original query, $T$ the set of available tools/skills, and $S_t$ the full history of executed steps and observations. This ensures the plan is always optimized with respect to the latest information (as in GoalAct (Chen et al., 23 Apr 2025)).
• Hierarchical and Local Planning: Hierarchical frameworks restrict the planning policy to select from a finite set of skills (e.g., Searching, Coding, Writing), delegating the instantiation of concrete tool invocations or low-level commands to a dedicated executor. Task-Decoupled Planning further localizes replanning to individual nodes of a sub-task DAG, isolating local failures and reducing computational overhead (Li et al., 12 Jan 2026).

Execution Modality

• Reactive Execution and Behavior Trees: Techniques such as the transformation of PDDL plans into behavior trees (BTs) optimize for parallelism and robustness by encoding causal relationships between actions and supporting dynamic recovery at run time (Martín et al., 2021). BTs allow action nodes to be ticked in sequences, parallel, or fallback structures, supporting fine-grained execution control.
• Skill Execution and Parametric Policies: In open-world or LLM-agent settings, a hierarchical executor routes plan steps to skill modules, whose implementation may rely on unified vision-language-action networks, tool calls, or code generation (e.g., emitting Python code for composite actions) (Mao et al., 2024, Chen et al., 23 Apr 2025).
• Predicate and Precondition Grounding: Robust execution in unstructured environments often employs predicate grounding, where LLMs generate preconditions for actions, which are then verified by perception modules before execution, pruning infeasible actions and facilitating recovery on failure (Rivera et al., 2024).

3. Interaction Between Planning and Execution

The integration of planning and execution in modern systems routinely adopts tight feedback loops:

• Plan–Execute–Observe–Replan Cycle: A prototypical loop iterates: plan update $\rightarrow$ select skill $\rightarrow$ hierarchical execution $\rightarrow$ observe result $\rightarrow$ append to history $\rightarrow$ repeat or terminate. Observations feed directly into the planner, triggering full or partial re-optimization of subsequent plan steps (Chen et al., 23 Apr 2025).
• Interactive and Adaptive Replanning: In interactive settings, such as robotics or user-driven visual analytics, replanning is triggered by new user intent, execution failures, environmental changes, or detected opportunity states. The system adapts the plan while preserving completed steps and minimizing disruption (Li et al., 2023, Borrajo et al., 2024).
• Event-Driven Replanning in Multi-Agent Systems: In architectures like CoMuRoS, event relevance is detected onboard each robot, and relevant task or environment changes are communicated to the central deliberative Task Manager for dynamic global replanning and reallocation of subtasks (Borate et al., 27 Nov 2025).

4. Empirical Benchmarks and Performance Analysis

The efficacy of task planning and execution frameworks is quantitatively benchmarked on domain-specific evaluation suites:

• LegalAgentBench (Chen et al., 23 Apr 2025): 300 legal reasoning and document drafting tasks; GoalAct attains up to +12.22% average improvement in success rate over baselines.
• VirtualHome (Arora et al., 4 Feb 2025): Anticipate & Act demonstrates 31% reduction in execution time and 12% reduction in plan length by jointly planning current and anticipated tasks.
• AI2-Thor and Embodied Task Suites (Rivera et al., 2024): ConceptAgent outperforms ReAct- and Tree-of-Thought–based LLM reasoning by up to 19% in task completion via predicate grounding and MCTS search.
• Prompt Engineering in Service Robotics (Bode et al., 2024): Adaptive Functions plus Example-in-Prompt achieves near-100% task completion in Fetch scenarios with significant reductions in execution latency for GPT-4 models.

Representative Results Table

Method	Domain	Key Gain
GoalAct	LegalAgentBench	+12.22% average success over baseline
Anticipate & Act	VirtualHome	–31% execution time, –12% plan length
ConceptAgent	AI2-Thor	×2–3 task completion rate vs. baseline
Skill-centric	RoboMatrix	50% improvement generalization on novel tasks (Mao et al., 2024)

5. Limitations and Open Challenges

Several fundamental and practical limitations remain across contemporary approaches:

• Granularity of Planning: Fixed-granularity decomposition can be inefficient; excessively coarse or fine sub-tasks impede adaptability and efficiency (Li et al., 12 Jan 2026).
• Skill and Action Space Limitation: Skill sets are often fixed and may require manual engineering or extension to adapt to new domains or modalities. Automated discovery and meta-learning of new compositional skills are recognized challenges (Chen et al., 23 Apr 2025, Mao et al., 2024).
• Dependency on LLM Quality and Perception Modules: Frameworks depending on LLMs are sensitive to model hallucinations, context window size, and network latency, with failures in predicate grounding propagating to infeasible executions (Rivera et al., 2024, Li et al., 2023).
• Handling Uncertainty and Partial Observability: Classical planners assume deterministic dynamics and full observability, while real-world environments necessitate robust strategies for probabilistic effects and incomplete information (e.g., via behaviors in TAMPER (Pan et al., 2024)).
• Scalability: Complexity grows with the number of sub-goals, skill types, or agents, and computational trade-offs arise when global planners are invoked at every step or in highly nested settings.

6. Future Directions

Key avenues suggested by recent research include:

• Learning and Adapting Planning Policies: Refinement of update policies via reinforcement learning or human feedback to yield robust global plans over diverse domains (Chen et al., 23 Apr 2025).
• Automated Skill Discovery and Meta-Learning: Developing methods for automatic extraction and incorporation of new skills, such as formal proof, multimodal perception, or situated manipulation, either via meta-learning or unsupervised segmentation (Chen et al., 23 Apr 2025, Mao et al., 2024).
• Tighter Integration with Reflection Modules: Integrating memory, reflection, or failure summarization modules to enhance long-horizon robustness and stability (Chen et al., 23 Apr 2025).
• Human-in-the-Loop and Multi-Agent Coordination: Extending architectures for robust multi-agent collaboration, explicit task-sharing, and seamless fallback to human execution or assistance in heterogeneous teams (Borate et al., 27 Nov 2025).
• Efficient Replanning and Error Isolation: Leveraging decoupled planning and execution scopes (e.g., TDP, BTs) to confine replanning to affected sub-tasks or plan segments, thus improving robustness and efficiency in complex or real-time scenarios (Li et al., 12 Jan 2026, Martín et al., 2021).

7. Comparative Summary and Outlook

Modern task planning and execution paradigms, spanning from global plan–skill decoupling (GoalAct, TDP) to skill-centric, hierarchical, and event-driven frameworks (CoMuRoS, RoboMatrix), represent a shift towards architectures that fuse abstraction, real-time perception, and robust execution. Empirical studies across diverse domains—legal reasoning, mobile service robotics, manufacturing, and visual analytics—demonstrate substantial gains in success rate, efficiency, and adaptability, but further progress depends on deeper integration of learning, compositionality, and human-AI collaboration (Chen et al., 23 Apr 2025, Mao et al., 2024, Borate et al., 27 Nov 2025, Li et al., 12 Jan 2026).