Subtask List (STL) in Incremental Planning

• Subtask List (STL) is a method that converts complex Signal Temporal Logic planning into sequential atomic tasks characterized by reachability and invariance constraints.
• It exposes nested temporal structures by rewriting bounded STL goals with symbolic witness times, reducing problem complexity and shortening effective planning horizons.
• Incremental solving of these subtasks improves runtime performance and safety guarantees in both linear and nonlinear systems, mitigating computational bottlenecks.

Searching arXiv for the target paper and closely related STL decomposition/planning work to ground the article. "Safe Planning through Incremental Decomposition of Signal Temporal Logic Specifications" studies trajectory planning for continuous-state cyber-physical systems from bounded Signal Temporal Logic (STL) goals, with the central objective of replacing one large, hard STL planning problem by a sequence of smaller subproblems that can be solved incrementally (Kapoor et al., 2024). The method targets the standard setting of motion and trajectory synthesis: given system dynamics, actuation limits, an STL formula over regions and obstacles, and a finite horizon, compute a trajectory that satisfies the formula. Its central claim is that the hidden sequential structure of nested temporal formulas can be exposed by rewriting them into primitive reachability and invariance constraints, scheduling those constraints online, and solving a series of short-horizon atomic tasks rather than one monolithic STL optimization (Kapoor et al., 2024).

1. Problem formulation and sources of complexity

The paper is motivated by a well-known scalability difficulty in STL planning. Direct “monolithic” STL planning, especially via mixed-integer encodings, scales poorly with both the specification horizon and the nesting depth of temporal operators. In existing encodings, each atomic proposition at each time step introduces binary variables, and nested operators require additional variables to represent temporal dependencies across time. The paper illustrates the issue with

$$\Phi = \phi_1 \land \phi_2,\qquad \phi_1 = \Diamond_{[0,10]} \bigl(R_1 \land \Diamond_{[20,30]} \Box_{[0,10]}(R_4)\bigr),\qquad \phi_2 = \Box_{[0,60]}(\neg R_3),$$

where the subformula $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$ already requires many binary variables under mixed-integer programming, and complexity grows exponentially with nesting depth (Kapoor et al., 2024).

The paper also emphasizes runtime issues that are not purely combinatorial. If the STL horizon exceeds the online planning horizon, receding-horizon solvers can return violating or poor plans. Limited online compute worsens this, and long-horizon planning is more sensitive to model mismatch because errors compound over longer predicted executions. The proposed decomposition is therefore presented not only as an encoding simplification but also as a way to reduce effective lookahead and to replan around realized satisfaction times rather than committing to one long open-loop solution (Kapoor et al., 2024).

This places the work within a broader STL planning landscape. Earlier decomposition work for multi-agent teams decomposes global STL formulas into local subtasks under structural assumptions (Charitidou et al., 2021), while runtime monitoring systems such as RTAMT focus on online evaluation of STL robustness rather than trajectory synthesis (Nickovic et al., 2020). The present method instead addresses online trajectory synthesis itself by incrementally decomposing bounded STL goals into schedulable local planning problems (Kapoor et al., 2024).

2. STL fragment and semantics

The paper recalls the general STL syntax

$$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$

with predicate $\mu \equiv \mu(\mathbf s(t))>0$, and uses the derived operators

$$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$

The actual decomposition method, however, is restricted to the fragment

$$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$

Thus propositional disjunction is allowed only inside non-temporal state formulas $p$, while temporal disjunction is not handled at the decomposition level; the lack of support for such disjunction is listed explicitly as a limitation (Kapoor et al., 2024).

The Boolean semantics are the standard finite-horizon ones: $s_t\models \mu \iff \mu(s(t))>0,$

$$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$

$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$0

The paper also recalls robustness semantics,

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$1

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$2

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$3

but these robustness semantics are used mainly for evaluation rather than for defining the decomposition (Kapoor et al., 2024).

A common misconception is that the method decomposes full STL because the preliminaries state syntax with $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$4. The paper is explicit that the decomposition itself is built only for $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$5, $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$6, conjunction, and propositional formulas; there is no direct decomposition rule for $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$7, and temporal disjunction is not structurally decomposed (Kapoor et al., 2024).

3. Primitive constraint representation and flattening

The key abstraction is that any formula in the supported fragment can be represented by two kinds of primitive task constraints: a reachability constraint, which ensures proposition $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$8 holds at least once in an interval $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$9, and an invariance constraint, which ensures proposition $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$0 holds at every time in $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$1. Formally, the paper defines a reachability set $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$2 and invariance set $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$3, each containing tuples $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$4, together with a set $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$5 of symbolic time-variable intervals $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$6. Symbolic times serve as existential witness times introduced by nested eventualities and are the mechanism by which sequential structure becomes explicit (Kapoor et al., 2024).

The recursive flattening procedure is

$$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$7

For an atomic propositional formula $$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$8,

$$\phi := \mu \mid \neg \phi \mid \phi \land \psi \mid \phi \lor \psi \mid \phi\,\mathcal U_{[a,b]}\,\psi,$$9

Thus a current-point state condition is represented as a degenerate reachability task. For conjunction,

$\mu \equiv \mu(\mathbf s(t))>0$0

flattening takes unions: $\mu \equiv \mu(\mathbf s(t))>0$1 Conjunction therefore means that all subtasks from both branches must be completed (Kapoor et al., 2024).

For eventuality,

$\mu \equiv \mu(\mathbf s(t))>0$2

flattening introduces a fresh symbolic time $\mu \equiv \mu(\mathbf s(t))>0$3 and shifts all descendant constraints by $\mu \equiv \mu(\mathbf s(t))>0$4: $\mu \equiv \mu(\mathbf s(t))>0$5

$\mu \equiv \mu(\mathbf s(t))>0$6

$\mu \equiv \mu(\mathbf s(t))>0$7

This is the central representation device for nested eventualities. For example, $\mu \equiv \mu(\mathbf s(t))>0$8 becomes one reachability subtask $\mu \equiv \mu(\mathbf s(t))>0$9 plus the interval constraint $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$0. For

$$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$1

the decomposition contains $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$2, $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$3, and symbolic bounds $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$4. This is precisely how an ordered subgoal structure emerges from nested $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$5-formulas (Kapoor et al., 2024).

For always,

$$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$6

flattening distinguishes reachability and invariance descendants. If $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$7 contributes reachability $$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$8, then

$$F_{[a,b]}\phi := \top\,\mathcal U_{[a,b]}\,\phi,\qquad G_{[a,b]}\phi := \neg F_{[a,b]}\neg\phi.$$9

If $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$0 contributes invariance $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$1, then

$$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$2

This matches identities such as

$$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$3

and captures the interpretation of formulas such as $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$4 as repeated future reachability obligations (Kapoor et al., 2024).

4. Symbolic time resolution, scheduling, and atomic tasks

Flattening leaves symbolic times unresolved, so the next stage is symbolic time resolution (STR). STR processes symbolic variables bottom-up and partially concretizes them into schedulable subtasks. Two rewrite rules are central. For a reachability constraint $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$5 containing symbolic $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$6, corresponding to nested $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$7, STR applies

$$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$8

This collapses nested eventuality windows into one concrete reachability interval (Kapoor et al., 2024).

For an invariance constraint $$\phi ::= F_{[a,b]}(\phi)\mid G_{[a,b]}(\phi)\mid \phi_1\land \phi_2\mid p, \qquad p ::= p_1\land p_2\mid p_1\lor p_2\mid \neg p\mid \mu.$$9 containing symbolic $p$0, corresponding to $p$1, STR interprets the formula as a reach-and-stay task. The paper splits it into two subtasks: first a reachability obligation to reach a state satisfying $p$2 sometime in $p$3, then an invariance obligation to remain in $p$4 for $p$5 more steps after the actual satisfaction time $p$6. This pattern is the main decomposition of nested $p$7-then-$p$8 formulas into sequential local goals (Kapoor et al., 2024).

After STR, the method computes a partial order $p$9 over constraints. A constraint $s_t\models \mu \iff \mu(s(t))>0,$0 if the time of satisfying $s_t\models \mu \iff \mu(s(t))>0,$1 must precede the time of satisfying $s_t\models \mu \iff \mu(s(t))>0,$2, using the interval relations in $s_t\models \mu \iff \mu(s(t))>0,$3, $s_t\models \mu \iff \mu(s(t))>0,$4, and the updated symbolic-time constraints. Thus a nested formula such as

$s_t\models \mu \iff \mu(s(t))>0,$5

becomes an explicit order among subgoals: reach $s_t\models \mu \iff \mu(s(t))>0,$6, then within shifted bounds reach $s_t\models \mu \iff \mu(s(t))>0,$7, then within shifted bounds reach $s_t\models \mu \iff \mu(s(t))>0,$8 (Kapoor et al., 2024).

The scheduler outputs a sequence of atomic tasks

$s_t\models \mu \iff \mu(s(t))>0,$9

where each atomic task has the non-nested form

$$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$0

with quantifier-free propositions $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$1, and different atomic tasks are arranged so their intervals do not overlap. To obtain such tasks, the scheduler performs slicing: when active constraints have mismatched intervals, it partitions the timeline into slices over which the active constraint set is constant and combines all active reachability propositions into one $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$2-part and all active invariance propositions into one $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$3-part. For example, if the active obligations are $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$4 and $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$5, slicing yields intervals $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$6, $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$7, $$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$8, and atomic tasks such as

$$s_t\models \neg\varphi \iff \neg(s_t\models\varphi), \qquad s_t\models \varphi_1\land \varphi_2 \iff (s_t\models\varphi_1)\land(s_t\models\varphi_2),$$9

(Kapoor et al., 2024).

This decomposition is closely related in spirit to other STL localization strategies, but its distinctive feature is the explicit runtime schedule induced by symbolic witness times rather than a static offline partition alone (Charitidou et al., 2021).

5. Incremental runtime planning and safety preservation

Runtime planning is incremental. The algorithm repeatedly selects the next currently enabled constraints according to the partial order, slices them into one or more atomic tasks, calls an external planner on one atomic task, appends the returned trajectory segment to the overall plan, extracts actual satisfaction times from that segment, and uses those realized times to instantiate symbolic variables and release successor subtasks (Kapoor et al., 2024).

The worked example in the paper makes the mechanism concrete. For

$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$0

the schedule can be written as

$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$1

$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$2

$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$3

$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$4

Operationally, the system first reaches $$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$5 while maintaining obstacle avoidance, then schedules the $$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$6 reachability window relative to the realized time $$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$7, then schedules the stay-in-$$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$8 task relative to the realized arrival time $$s_t\models F_{[a,b]}(\varphi) \iff \exists t'\in[t+a,t+b]\;\text{s.t.}\; s_{t'}\models \varphi,$$9, while continuously preserving $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$00 across overlapping slices (Kapoor et al., 2024).

The paper argues that this reformulation yields three practical benefits. First, each online solve no longer contains nested temporal operators, which reduces binary-variable blowup in mixed-integer encodings. Second, the effective horizon per solve is shorter, reducing runtime and improving feasibility under limited compute. Third, because replanning occurs after each achieved subtask using realized satisfaction times, the planner is less exposed to compounded model error than long open-loop optimization over the full composite formula (Kapoor et al., 2024).

The method is planner-agnostic. It does not introduce a new low-level optimal control formulation; each atomic task is passed to an existing STL-capable planner. In implementation, the paper uses stlpy together with Drake, solving with Gurobi for mixed-integer convex programming in linear systems and SNOPT for nonlinear systems. The experimental systems are a planar double integrator and a planar unicycle (Kapoor et al., 2024).

6. Guarantees, empirical performance, and limitations

The method is accompanied by soundness results. First, flattening is sound: if all flattened constraints are satisfied under an instantiation of symbolic times consistent with $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$01, then the original formula is satisfied. Second, STR is sound: satisfaction of resolved constraints implies satisfaction of the original flattened constraints. Third, the scheduler ensures the synthesized plan satisfies all scheduled reachability and invariance constraints under the time assignments discovered online. The paper summarizes the final guarantee as

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$02

The guarantee is therefore soundness, or correct-by-construction satisfaction, provided the underlying atomic-task planner returns trajectories satisfying each atomic formula (Kapoor et al., 2024).

The approach is not complete. Incremental decomposition may miss globally feasible or better solutions, and the paper explicitly notes that it prioritizes satisfaction over global optimality. The decomposition level also does not support temporal disjunction, and there is no direct rule for $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$03. A further practical limitation is that for depth $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$04 or very simple formulas, the overhead of flattening and scheduling can outweigh the benefit (Kapoor et al., 2024).

The empirical evaluation uses motion-planning benchmarks built from common patterns Reach, Avoid, Sequenced Visit, and Stabilization. Representative formulas include

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$05

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$06

$\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$07

Against standard MICP and reduced MICP, the main finding is that StlInc is comparable or slightly slower on shallow formulas but much better on nested ones. In the linear environment, for $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$08 with horizon $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$09 and nesting depth $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$10, both baselines time out at 30 minutes, while StlInc solves in $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$11 s with positive robustness $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$12. For $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$13, StlInc solves in $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$14 s versus $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$15 s and $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$16 s for the baselines. In the nonlinear environment, $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$17 solves in $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$18 s versus $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$19 s for reduced MICP, and $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$20 in $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$21 s versus $\Diamond_{[20,30]}\Box_{[0,10]}(R_4)$22 s, with the baseline failing to find a satisfying plan (Kapoor et al., 2024).

Within the wider STL literature, this positions the method between monolithic synthesis and task-level decomposition. Runtime monitors such as RTAMT evaluate STL or IA-STL online and can quantify robustness, but they do not perform the incremental synthesis described here (Nickovic et al., 2020). Multi-robot STL methods have also separated assignment from low-level motion by planning over path choices or swarm abstractions (Liu et al., 4 Mar 2025, Cheng et al., 17 Jun 2025). A plausible implication is that the incremental decomposition strategy of (Kapoor et al., 2024) is particularly attractive when the main bottleneck is nested temporal structure in a single-agent or low-dimensional continuous planning problem, whereas other abstractions become more natural when coordination or large team size dominates.

The paper’s methodological takeaway is precise: nested bounded STL specifications can be converted into a dependency-aware list of reachability and invariance subtasks, symbolic witness times can be instantiated online as earlier subtasks are completed, and the resulting atomic tasks can be solved sequentially so that satisfaction of the scheduled list implies satisfaction of the original STL specification (Kapoor et al., 2024).