Task-Hierarchical Bilevel MPC

• Task-Hierarchical Bilevel MPC is a control paradigm where a higher layer selects task-level decisions while a lower layer computes optimal trajectories over a receding horizon.
• It encompasses approaches from formal bilevel MPC to hierarchical cascades, enabling lexicographic task priorities, contact planning, and human-aware interactions in robotics.
• Practical applications in mobile manipulation and crowd navigation demonstrate improved coordination, computational efficiency, and robust stability guarantees compared to single-layer MPC.

Searching arXiv for papers on task-hierarchical and bilevel MPC to ground the article. Searching arXiv for exact papers cited in the source material. Task-Hierarchical Bilevel Model Predictive Control denotes a family of predictive-control architectures in which a higher layer selects task-level variables—such as reference sequences, strict task priorities, contact decisions, or coordination signals—while a lower layer computes dynamically feasible trajectories, control inputs, or modeled responses over a receding horizon. In the strict optimization-theoretic sense, the upper problem is constrained by the optimal solution map of a lower predictive controller; in broader robotics usage, the term also covers hierarchical predictive cascades that are strongly bilevel-like in control structure even when lower-level optimality conditions are not embedded explicitly (Moriyasu et al., 31 Mar 2026, Rigo et al., 2022).

1. Conceptual scope and taxonomy

The literature uses the term across several distinct, partially overlapping constructions. The cleanest distinction is between formal bilevel MPC, where lower-level optimality appears inside the upper problem, and hierarchical predictive control, where the upper layer supplies references, targets, or reduced-order decisions to a lower MPC without forming a single nested mathematical program. A second distinction separates task hierarchy in the robotics sense—strict priority among robot objectives—from agent hierarchy or leader–follower hierarchy, where the lower level represents the response of humans, followers, or subordinate controllers rather than lower-priority robot tasks.

Formulation family	Representative papers	Characterization
Reference-sequence bilevel MPC	(Moriyasu et al., 31 Mar 2026)	Upper layer selects a reference sequence; lower linear MPC tracks it
Lexicographic task-hierarchical MPC	(D'Orazio et al., 21 Nov 2025, Du et al., 10 Mar 2026)	Multiple robot tasks ordered by strict priority through lexicographic optimization
Cascade or bilevel-like hierarchical MPC	(Rigo et al., 2022, Lee et al., 2022, Lee et al., 2020)	Upper layer supplies contact plans or nominal task motion; lower MPC realizes them
Agent-response or Stackelberg bilevel MPC	(Samavi et al., 2023, Thirugnanam et al., 7 Feb 2025)	Lower level models human or follower optimal response
Learned hierarchical predictive architectures	(Chen et al., 11 Feb 2026, Li et al., 12 Feb 2025, Peng et al., 6 Feb 2025)	Hierarchical prediction, symbolic planning, or RL-guided subgoal generation rather than classical online bilevel MPC

This taxonomy matters because many papers are hierarchical without being formally bilevel, and several papers are genuinely bilevel without being task-hierarchical in the classical task-priority sense.

2. Formal optimization structures

A canonical task-hierarchical bilevel MPC for linear systems is given by an upper layer that chooses a stacked reference sequence $\Theta$ and a lower-level tracking MPC that chooses the control sequence $U$. In condensed form, the bilevel problem is

$$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$

subject to

$$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$

For this class, a smooth single-level reduction is available under the verifiable nonsingularity of the block matrix $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$. The reduced problem replaces lower-level optimality by the affine stationarity relation

$$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$

and, in the convex case, the solution is unique and equivalent to a corresponding centralized MPC, which allows inheritance of closed-loop properties such as stability (Moriyasu et al., 31 Mar 2026).

A different formal route is the KKT-based embedding of lower-level optimality. In SICNav, the upper robot MPC optimizes robot motion while each human is constrained to solve a relaxed ORCA problem; lower-level optimality is enforced through stationarity, complementarity, primal feasibility, and dual feasibility,

$$\nabla_{v,\xi}\mathcal L = 0,\qquad \lambda^\top g=0,\qquad g\le 0,\qquad \lambda\ge 0,$$

yielding an MPCC after reformulation. SM$^2$ITH adopts the same mathematical pattern for interactive human prediction, but combines it with a lexicographic stack of mobile-manipulation task objectives and discrete-time control barrier functions (Samavi et al., 2023, D'Orazio et al., 21 Nov 2025).

A third formal structure replaces explicit lower-level optimality constraints by incentive-mediated implementability. In hierarchical convex MPC viewed as a Stackelberg game, the upper-level BiMPC augments the follower cost by a linear incentive $\langle \lambda,w\rangle$. For a single strongly convex lower-level MPC, the incentivized bilevel problem is exactly equivalent to a team-optimal single-level problem, and an implementing incentive satisfies the subgradient inclusion

$$-\lambda^\star \in \partial_w g(w^\star;\xi_0,u^\star).$$

For multiple lower-level MPCs sharing one incentive, exact controllability is replaced by a bounded implementation error

$U$0

which leads to a robust single-level reformulation (Thirugnanam et al., 7 Feb 2025).

3. Mechanisms for enforcing task hierarchy

Strict task hierarchy is usually expressed lexicographically. In HTMPC for mobile manipulators, the controller solves

$U$1

over a horizon subject to robot dynamics, admissible state and input sets, initial conditions, and self-collision avoidance through a signed-distance constraint. The lexicographic problem is implemented as a sequence of single-task MPC solves, and lower-priority tasks are prevented from degrading higher-priority tasks by componentwise non-degradation constraints of the form

$U$2

This produces a strict hierarchy without collapsing tasks into a weighted sum (D'Orazio et al., 21 Nov 2025).

For sequential mobile manipulation, lexicographic hierarchy is extended directly to task lists supplied by a planner. The controller receives a sliding ordered sublist $U$3 and solves a sequence of nonlinear predictive problems with costs $U$4. Higher-priority solutions are preserved through lexicographic constraints $U$5, and a relaxed implementation permits $U$6 during line search. The paper further proposes a decoupled “box” formulation

$U$7

as an inner approximation that improves convergence in practice (Du et al., 10 Mar 2026).

Several influential architectures are hierarchical in control logic but not strict bilevel programs. In non-prehensile loco-manipulation, two cascading MPCs split the problem into an upper contact optimizer over the interaction variables $U$8 and a lower locomotion-aware MPC that realizes the desired force and contact point through body motion and foot reaction forces. The hierarchy is strongly bilevel-like because the upper layer decides how and where to push, while the lower layer realizes those decisions under robot dynamics and gait constraints, but the lower MPC is not embedded as a KKT system inside the upper problem (Rigo et al., 2022).

A comparable layered construction appears in real-time manipulator MPC with singularity-tolerant hierarchical task control. There, a hierarchical inverse-kinematics or inverse-dynamics controller first produces a nominal task-consistent trajectory using recursive null-space projections, and MPC then linearizes around that nominal motion and solves a fast QP to enforce joint, velocity, and torque limits. Task hierarchy is therefore encoded upstream in the nominal generator rather than inside the receding-horizon optimization itself (Lee et al., 2022).

A more direct single-level surrogate of hierarchical control is the QCQP formulation for underactuated and constrained robots. Instead of nested task optimizations, the controller embeds task hierarchy by quadratic inequalities comparing relative task errors over the horizon, thereby approximating strict hierarchy inside one convex predictive program rather than a formal bilevel or multilevel optimization (Lee et al., 2020).

4. Representative robotic domains and empirical behavior

The most mature task-hierarchical bilevel formulations are now found in mobile manipulation and human-aware robotics. SM$U$9ITH combines lexicographic HTMPC with interactive human prediction and is validated on Stretch 3 and Ridgeback–UR10 across delivery tasks with different navigation and manipulation priorities, sequential pick-and-place tasks with different human models, and adversarial human interaction. The paper reports that interactive prediction improves coordination relative to weighted objectives, constant-velocity motion models, and purely reactive baselines, with the effect especially pronounced on the more agile Ridgeback–UR10 platform (D'Orazio et al., 21 Nov 2025).

In contact-rich locomotion-manipulation, hierarchical MPC has been used to optimize both contact force and contact location during pushing with a quadruped. The architecture tracks straight, curved, velocity-changing, and obstacle-avoidance-style object trajectories, and the paper notes a practical speed distinction: straight pushing can track speeds greater than $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$0, whereas sharp turns require reducing speed to around $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$1 (Rigo et al., 2022).

For sequential mobile manipulation, lexicographic HTMPC has been demonstrated on a 9-DoF Ridgeback–UR10 platform. Relative to task-prioritized inverse differential kinematic control, the method improves hierarchical trajectory tracking performance by $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$2 on average when facing task changes, robot singularity, and reference variations; relative to a single-task architecture, it yields a shorter task-space path and an execution time $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$3 times faster in a sequence of delivery tasks (Du et al., 10 Mar 2026).

At the manipulator level, real-time hierarchical-MPC tracking around singularities and torque limits has been validated on industrial robots, with the resulting QP-based implementation achieving an update frequency higher than $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$4 kHz and outperforming an operational-space-control baseline in tracking accuracy. The reported scenarios include carrying heavy payloads while accounting for torque limits and end-effector control while avoiding singularities (Lee et al., 2022).

In crowd navigation, the hierarchy is over interacting agents rather than prioritized robot tasks, but the empirical behavior is still instructive for bilevel MPC. SICNav reported success $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$5 with average navigation time $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$6 s for $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$7 humans and success $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$8 with average navigation time $$\mathbf P_1:\quad \min_{\Theta,U} F_u(U;x_0) \quad\text{s.t.}\quad G_u(U;x_0)\le 0,$$9 s for $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$0 humans, outperforming a decoupled constant-velocity-prediction baseline in navigation time, collision frequency, and freezing frequency (Samavi et al., 2023).

5. Computation, feasibility, and closed-loop properties

The principal computational challenge is that nested predictive optimization often produces MPCC- or MPEC-like structure. KKT-based formulations inherit complementarity, nonconvexity, and possible constraint-qualification pathologies. SICNav explicitly identifies the resulting reformulation as an MPCC and reports solution times between $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$1 and $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$2 in the current Python implementation, with variable dimension scaling as $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$3 because of pairwise crowd interactions (Samavi et al., 2023).

The main tractability breakthrough for formally bilevel tracking MPC is the smooth reduction for linear systems. In the quadrotor experiments comparing hierarchical and centralized formulations, average solve times were $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$4 ms for $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$5, $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$6 ms for the original bilevel formulation $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$7, $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$8 ms for the reduced formulation $$U=\arg\min_V \left\{F_l(V;\Theta,x_0)\mid G_l(V;x_0)\le 0\right\}.$$9, and $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$0 ms for the centralized MPC $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$1. Under the nonsingularity of $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$2, $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$3 is exact in optimal value, and under additional convexity assumptions it realizes the same unique control input as the centralized MPC, so stability can be inherited from the centralized design (Moriyasu et al., 31 Mar 2026).

Hierarchical task controllers that are not formal bilevel programs often trade exact nested optimality for online practicality. The sequential mobile-manipulation HTMPC runs with $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$4, $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$5, and $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$6, with average compute time $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$7 ms and control frequency $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$8 Hz; the lower-rate predictive layer is compensated by warm starts, line-search tolerances, and barrier relaxations on state, control, and collision constraints (Du et al., 10 Mar 2026).

Feasibility and safety guarantees vary sharply across formulations. In learned hierarchical predictive control, a safe invariant set $$\bar\Gamma=\bar B^\top\bar Q\bar Z_x+\bar R\bar Z_u$$9 together with a backup controller $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$0 is sufficient to prove feasible execution of a new task in a new environment, provided the initial condition lies in $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$1. The proof uses the lifted target-set construction

$$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$2

and a shifting-horizon fallback argument (Vallon et al., 2021).

For multi-agent capacity-constrained systems, hierarchical learning is coordinated across levels through high-level safe sets, low-level edge-specific safe sets, and monotone updates of learned edge-capacity usage. The resulting architecture is iteratively feasible, and the number of unique non-depot nodes visited satisfies

$$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$3

so performance is non-decreasing across iterations (Vallon et al., 2024).

The main limitations recur across the literature. Formal exactness usually requires strong convexity, convex feasible sets, or verifiable nonsingularity conditions; incentive-based coordination additionally assumes convex lower-level MPCs and, in the multi-follower theory, a shared-domain structure that excludes heterogeneous state constraints across followers (Thirugnanam et al., 7 Feb 2025). Safety guarantees are typically model-based and conditional on prediction fidelity, state estimation, and online feasibility; SM$$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$4ITH makes this explicit for DT-CBF-based human-robot safety (D'Orazio et al., 21 Nov 2025). In cascade architectures, upper-level feasibility is often only indirect because lower-level realizability is not embedded in the upper problem, as in contact-optimization MPC for quadruped pushing (Rigo et al., 2022).

6. Learned, symbolic, and world-model extensions

Recent work has extended the task-hierarchical idea beyond classical optimization by replacing the upper level with learned symbolic or perceptual predictors. H-WM is a hierarchical world model with a logical world model $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$5, a visual world model, and a sub-goal VLA policy. It is explicitly described as conceptually close to task-hierarchical bilevel MPC but not a classical bilevel MPC algorithm, because execution is driven by learned prediction and subtask-level replanning rather than an online constrained upper–lower optimization. On LIBERO-LoHo, the H-WM-guided policy achieves $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$6 QS/SR, compared with $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$7 for logic-guided and $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$8 for language-guided control (Chen et al., 11 Feb 2026).

At the symbolic-planning interface, IVNTR learns neural predicates for bilevel planning directly from demonstrations. The method is not an MPC formulation, but it is highly relevant as an abstraction-learning front end for upper task layers, because it learns the symbolic state abstractions and effect structures that a task-hierarchical MPC could use. Across six simulated domains, IVNTR attains an average unseen-task success around $$(\bar B^\top\bar Q\bar B+\bar R)U-\bar\Gamma\Theta+\bar B^\top\bar Q\bar A x_0=0,$$9 and demonstrates real-world mobile-manipulation generalization to new objects, new states, and longer horizons (Li et al., 12 Feb 2025).

A related but distinct direction combines learned high-level behavior with low-level MPC. BiM-ACPPO uses an RL policy to choose intermediate decision variables—waypoint, reference velocity, and lane-change action—which are decoded into an intermediate reference state $$\nabla_{v,\xi}\mathcal L = 0,\qquad \lambda^\top g=0,\qquad g\le 0,\qquad \lambda\ge 0,$$0 for a nonlinear MPC controller. The paper is hierarchical RL plus MPC rather than classical bilevel control, and its “bilevel” terminology primarily refers to curriculum learning; nevertheless, it shows a practically important upper-to-lower interface in which the high-level layer sets the reference manifold for the lower predictive controller. In overtaking, the reported few-shot fine-tuned success rate is $$\nabla_{v,\xi}\mathcal L = 0,\qquad \lambda^\top g=0,\qquad g\le 0,\qquad \lambda\ge 0,$$1 (Peng et al., 6 Feb 2025).

These extensions suggest a widening interpretation of task-hierarchical bilevel MPC. The core optimization-based paradigm remains the coupling of task-level predictive decisions with a lower realization layer, but the upper level is increasingly instantiated by learned abstractions, symbolic models, or subgoal predictors rather than by a purely analytic optimizer. The conceptual boundary of the field therefore runs between nested optimality, which defines strict bilevel MPC, and hierarchical predictive interfacing, which defines the broader robotics family to which many influential recent systems belong.