Progress-Aware Symbolic Sub-Task Plans

• Progress-aware symbolic sub-task plans are formalized decompositions that break complex objectives into sequential sub-tasks with explicit progress tracking.
• They integrate hierarchical representations, formal verification methods, and adaptive feedback loops to ensure robust execution in dynamic environments.
• Applications include multi-robot coordination, human-aware task execution, and industrial assembly where real-time replanning and state validation are critical.

A progress-aware symbolic sub-task plan is a formalized sequence of symbolic actions or state transitions designed to achieve a complex objective through explicit sub-task decomposition, with mechanisms for monitoring and adapting progress at each step. In contemporary planning systems—especially those integrating neuro-symbolic and LLM reasoning—progress-awareness is realized through hierarchical representations, formal verification, and closed-loop feedback, enabling robustness to stochasticity, human interaction, or environmental changes.

1. Formal Representations for Progress-Aware Sub-Tasks

Progress-aware planning fundamentally relies on a symbolic decomposition of complex tasks into sub-tasks, which are systematically tracked and verified throughout execution. Prevalent formalizations include:

• Hierarchical Temporal Logic (HTL): Task specifications are defined as a hierarchy $L_1 \to L_2 \to \dots \to L_K$ of syntactically co-safe Linear Temporal Logic (sc-LTL) formulas. A root formula $\varphi_1^1$ encodes the task as a sequence and/or conjunction of sub-goals, recursively composed from child formulas at lower levels. Only the leaf formulas reference atomic propositions corresponding to concrete skills (e.g., Pick, Place, Handover). This induces a tree/graph that makes explicit the sequential, parallel, or conditional dependencies between sub-tasks (Hu et al., 10 Feb 2026).
• Explicit Sub-Goal Sequences: Complex tasks $T = (\mathcal S, s_0, g, \mathcal A)$ are decomposed into ordered macro-actions $(ma^1, ..., ma^m)$, each of which expands to a block of atomic actions $aa_j^i \in \mathcal A$. The result is a two-level plan structure: symbolic (macro) plan and expanded (atomic) plan (Cornelio et al., 6 Apr 2025).
• Symbolic Progress Predicates: In structured assembly, progress is formalized using sets of "installed" $I_t$ and "uninstalled" $U_t$ components, with progress measured as $|I_t|/|V|$ and the admissible action frontier $$T_t = \{ v_i \in U_t \,|\, Pre(v_i) \subseteq I_t \}$$ (Chen et al., 2 Jan 2026).
• MCTS with Prioritized Node Expansion: The planner tracks which sub-goals $G_{enc}(s)$ have been satisfied at each search node, dynamically prioritizing expansions that achieve new sub-goals and maintaining a progress-aware depth-first search (Pfeiffer et al., 2023).

2. Architectures and Algorithms for Progress-Awareness

Progress-aware symbolic planning integrates several computational modules:

• LLM-Grounded Specification Extraction: Natural language commands are parsed by an LLM into ordered sub-goals and structured as hierarchical symbolic specifications (e.g., JSON-encoded H-LTL), including parallel/sequential hints. These form the skeleton of the symbolic sub-task plan (Hu et al., 10 Feb 2026).
• Product Automata and Receding Horizon Planning: Each sub-task formula is mapped to a finite automaton. Robot transition systems are synchronized via a product automaton, enriched with in-specification and inter-specification switching transitions that allow for flexible multi-agent task allocation. The plan is solved using A* in a receding horizon loop, providing dynamic plan adaptation in response to environmental changes (Hu et al., 10 Feb 2026).
• Neuro-Symbolic Dual Memory: Progress Memory stores neural "anchors" (blueprints of successful sub-task trajectories) for semantic progression, while Feasibility Memory holds symbolic rules (e.g., Python checkers) to strictly filter candidate actions for logical consistency and executability. Actor policies balance scores from both memories to select each action (Wen et al., 3 Apr 2026).
• Symbolic State Tracking and Verification: At each executed step, the expected symbolic state (as predicted by the plan) is compared to observed world state, enabling immediate detection of drift or execution anomalies. A symbolic validator (e.g., PDDL-based) can detect unsatisfiable preconditions or unmet postconditions, triggering re-synthesis for the remaining suffix when progress falters (Cornelio et al., 6 Apr 2025).
• Iterative Correction and Contrastive Ranking: In frameworks like SymPlanner, actions invalid under the symbolic transition function are identified and replaced on the fly (IC), while completed plans are ranked by expected goal achievement and log-probabilities, ensuring that only progress-maximizing sequences are committed (Xiong et al., 2 May 2025).
• Minimal-Change Replanning: In collaborative assembly, a minimal plan-edit distance is maintained between consecutive plans during runtime, restricting updates to precisely those sub-tasks whose frontier admissibility or assignment has changed due to environmental events (e.g., unexpected human installations) (Chen et al., 2 Jan 2026).

3. Mechanisms for Monitoring, Feedback, and Reactive Adaptation

A hallmark of progress-aware plans is their ability to adaptively track and respond to execution:

• Progress-Triggered Re-Optimization: Upon completion of any sub-task (e.g., after crossing an NFA edge in the product automaton), the planner re-evaluates the remaining specification and re-plans, allowing agents to redistribute work dynamically and mitigate delays (Hu et al., 10 Feb 2026).
• Progress Metrics and Validation: Progress during execution is quantified via mathematical measures; for example, the overlap ratio $\varphi_1^1$0 between expected and observed symbolic states provides an explicit, thresholded signal for triggering replanning or correction (Cornelio et al., 6 Apr 2025).
• Drift Detection and Memory Re-Anchoring: When neural progress monitors register repeated failures to achieve a new anchor, the system extends retrieval windows or re-invokes the blueprint planner, realigning policy to a semantically valid sub-task trajectory (Wen et al., 3 Apr 2026).
• Reactive Safety and Predictive Adaptation: Robot teams respond instantly to kinematic infeasibility or predicted human intrusion by halting execution, resetting planning horizons, and incorporating real-time perception into the plan synthesis process. Predictive horizon adaptation precomputes alternate plans in anticipation of workspace changes, ensuring continuity (Hu et al., 10 Feb 2026).
• Frontier-Based Task Generation: In collaborative assembly, only frontier actions admissible under the current symbolic state are considered, preventing invalid execution and focusing all planning on the maximally-progressive set of next actions (Chen et al., 2 Jan 2026).

4. Empirical Performance and Theoretical Guarantees

Multiple systems establish rigorous correctness, optimality, and efficiency properties:

• Correctness by Construction: Paths produced by automata-based planners or validators are guaranteed, by formal semantics, to be consistent with the symbolic specification (e.g., sc-LTL hierarchy or PDDL domain) (Hu et al., 10 Feb 2026, Cornelio et al., 6 Apr 2025).
• Optimality in Cost: Weighted A* search over the combined team automaton yields team-optimal solutions with admissible heuristics, minimizing sum execution cost $\varphi_1^1$1 (Hu et al., 10 Feb 2026).
• Efficient Scaling: Hierarchical decomposition dramatically reduces automaton state space, and progress-prioritized MCTS provides linear complexity scaling in the number of sub-goals $\varphi_1^1$2 when bridging factor $\varphi_1^1$3 is properly set (Pfeiffer et al., 2023).
• Empirical Outcomes: Across ALFWorld, WebShop, and TextCraft benchmarks, dual-memory agents realize large increases in task success, with invalid action rates (IAR) dropping by up to $\varphi_1^1$4 and trajectory lengths (ATL) reduced by up to 36% (Wen et al., 3 Apr 2026). In collaborative robot experiments, solution rates exceed 93% even in 16-subtask scenarios, with real-world deployments maintaining ≥90% success under dynamic human interventions and minimizing plan instability (Hu et al., 10 Feb 2026, Chen et al., 2 Jan 2026).
• Progress and Verification Metrics: Plan correctness (PC), execution success (ES), plan length discrepancy (LD), and block-level verification rates provide quantitative probes of progress-aware planning efficacy and minimality (Cornelio et al., 6 Apr 2025).

5. Illustrative Case Studies and Domains

The progress-aware symbolic sub-task paradigm underlies a variety of recent system designs:

• Multi-Robot Human-Aware Task Execution: LLM-derived sc-LTL task trees are used to allocate, synchronize, and adaptively reallocate sub-tasks across robot teams in response to user movements and environmental events, as demonstrated in line-topology and dynamic collaborative scenarios (Hu et al., 10 Feb 2026).
• Hierarchical Robot Cooking and Service Tasks: Plans for preparing and serving items are decomposed into macro- and atomic actions via KG-RAG and symbolically verified with real-time corrections using observations of workspace state, e.g., regrasping if a glass slips (Cornelio et al., 6 Apr 2025).
• Block-Based Visual Programming and AI Education: Progressions of subtasks (minimally-different, well-ordered code fragments) are synthesized to assist neural program synthesizers and human learners, quantitatively demonstrating that gradual, progress-aligned task sequences improve learning efficiency and success (Tercan et al., 2023).
• Collaborative Structured Assembly: Symbolic tracking of installed components and minimally-editable task allocation plans yield robust, verifiable assembly in the presence of unpredictable human interventions and perception noise (Chen et al., 2 Jan 2026).
• Industrial Multi-Goal Robot Inspection: Progress-aware MCTS discovers very long valid symbolic action sequences (up to 145 steps/48 sub-goals) for collaborative manipulation and inspection, with scaling guarantees rooted in prioritized expansion and kinematic pruning (Pfeiffer et al., 2023).

6. Comparison of Principal Approaches

Approach	Progress Monitoring	Adaptation Trigger	Correctness Mechanism
HTL+RHP (Hu et al., 10 Feb 2026)	Hierarchical formula progress	NFA edge completion, safety/prediction failures	Automata-based synthesis
KG-RAG + Symbolic Validation (Cornelio et al., 6 Apr 2025)	State/graph overlap ($\varphi_1^1$5)	Shortfall in expected state, unsat preconds	PDDL/VAL validator
Dual Memory (Wen et al., 3 Apr 2026)	Neural anchor tracking	Drift detection, failure-induced rules	Dual-memory decision filter
Iterative Correction (Xiong et al., 2 May 2025)	Sub-goal milestone checks	Invalid transitions	PDDL transition simulator
Minimal-Change HRC (Chen et al., 2 Jan 2026)	Installed/uninstalled frontier	Human intervention, state mismatch	Rule-set reconciliation
MCTS+PNE (Pfeiffer et al., 2023)	Sub-goal set G_enc(s)	New sub-goal achieved, bridging exhaustion	UCT with prioritized expansion

These architectures collectively demonstrate that progress-aware symbolic sub-task planning is characterized by explicit symbolic modeling, continuous state/task verification, and adaptive feedback-driven synthesis, all of which are critical for robust, efficient, and correct long-horizon task execution in contemporary AI and robotics.

7. Open Challenges and Implications

While empirical results indicate strong gains over naïve LLM- or flat symbolic planning baselines, emergent challenges persist. Hierarchical sc-LTL or automata-based approaches, though vastly more scalable than flat LTL, may still face exponential blow-ups in dynamic open-world settings. Robustness to perceptual noise and casual user interventions demands both high-fidelity symbolic state estimation and rapid, controllable replanning. As systems broaden to new domains—ranging from industrial assembly to program synthesis in education—a central research direction is the co-design of symbolic progress models, neuro-symbolic validation modules, and adaptive feedback policies, aiming to unify correctness, tractability, and human-aware fluency (Hu et al., 10 Feb 2026, Cornelio et al., 6 Apr 2025, Tercan et al., 2023, Chen et al., 2 Jan 2026).