Trajectory-Adaptive Bi-level Optimization

• Trajectory-adaptive bi-level optimization is a framework that decomposes complex trajectory planning into nested upper-level and lower-level optimization tasks.
• It employs efficient gradient-based methods, such as KKT-based implicit differentiation, to adaptively update trajectory parameters and optimize both spatial and temporal dimensions.
• This approach has demonstrated real-time capabilities and scalability across UAVs, robotics, and autonomous vehicle applications while effectively handling high-dimensional constraints.

Trajectory-adaptive bi-level optimization refers to a class of bilevel optimization methods and formulations in which the variables governing a system’s trajectory—either in time, space, or configuration—are treated as adaptive quantities updated at one or both optimization levels. These frameworks exploit the inherent temporal or spatial structure of trajectory planning, control, or learning problems, yielding tractable, scalable, and often real-time-capable solutions to complex, constrained, and high-dimensional design tasks.

1. Conceptual Foundations and Formalism

In a general trajectory-adaptive bilevel (TABL) setup, the optimization is formulated as a nested program:

• The lower-level problem (LLP) computes, for fixed "trajectory parameters" (e.g., timing, waypoints, shape coefficients), the optimal system evolution or control sequence under constraints.
• The upper-level problem (ULP) adapts these trajectory parameters to further minimize some global cost or satisfy outer constraints, often using sensitivities (hypergradients) propagated through the solution map of the lower level.

Mathematically, for trajectory parameter $\zeta$ (e.g., segment durations, waypoints, policy weights): $$\min_{\zeta \in \mathcal{Z}^{UL}} F\left(\zeta, x^*(\zeta)\right), \quad \mathrm{where}\quad x^*(\zeta) = \arg\min_{x \in \mathcal{X}^{LL}} f(\zeta, x) \ \text{s.t.} \ g_{LL}(\zeta,x) \leq 0$$ Here, $F$ encodes the outer objective (e.g., total cost, constraint violation), and $f$ encodes the inner cost (e.g., minimum snap, minimum energy, trajectory feasibility).

TABL schemes crucially exploit trajectory structure in $\zeta$, employing parameterizations (piecewise-polynomial, B-spline, Bezier, basis-function, or policy representations) that are amenable to convex optimization, sensitivity analysis, or efficient heuristics in the LLP, and admit meaningful physical, geometric, or performance interpretation in the ULP.

2. Algorithmic Structures and Differentiation

Decoupled State-Timing and Spatial-Temporal Decomposition

In UAV trajectory planning, for example, one approach decouples spatial variables (trajectory shape) and timing variables (per-segment durations) (Sun et al., 2018). For fixed timing, the spatial LLP is a convex QP: $$\min_{c} c^\top P(y) c + w\mathbf{1}^\top y\quad \text{s.t.}\ G(y)c \leq h, \ H(y)c = m$$ with $y$ the timing vector. The ULP then optimizes $y$, updating timing allocations with analytic gradients: $$\nabla_y J^\star(y) = \nabla_y J(c^*, y) + (\lambda^{*\top} \nabla_y[G(y)c^* - h] + \nu^{*\top} \nabla_y[H(y)c^* - m])$$ This leverages Lagrange multipliers $(\lambda^*, \nu^*)$ extracted as NLP dual variables. Such sensitivity analysis enables efficient gradient-based updates (often projected) at the upper level.

Lifting of Constraints and Efficient KKT Differentiation

For UAV minimum-snap planning, safety polytope constraints (e.g., collision avoidance) are lifted from the LLP to the ULP by augmenting the LLP with extra equality constraints and moving convex inequalities to the ULP (Chen et al., 2022). The LLP then becomes a pure equality QP, solvable via KKT matrix factorization: $$\min_{\zeta \in \mathcal{Z}^{UL}} F\left(\zeta, x^*(\zeta)\right), \quad \mathrm{where}\quad x^*(\zeta) = \arg\min_{x \in \mathcal{X}^{LL}} f(\zeta, x) \ \text{s.t.} \ g_{LL}(\zeta,x) \leq 0$$0 KKT-based implicit differentiation, as implemented in differentiable QP solvers such as OptNet, yields the required $$\min_{\zeta \in \mathcal{Z}^{UL}} F\left(\zeta, x^*(\zeta)\right), \quad \mathrm{where}\quad x^*(\zeta) = \arg\min_{x \in \mathcal{X}^{LL}} f(\zeta, x) \ \text{s.t.} \ g_{LL}(\zeta,x) \leq 0$$1 and $$\min_{\zeta \in \mathcal{Z}^{UL}} F\left(\zeta, x^*(\zeta)\right), \quad \mathrm{where}\quad x^*(\zeta) = \arg\min_{x \in \mathcal{X}^{LL}} f(\zeta, x) \ \text{s.t.} \ g_{LL}(\zeta,x) \leq 0$$2 in a single backward pass. The projected gradient step at the ULP simultaneously updates both waypoints and timing splits, with feasibility projections ensuring satisfaction of spatial and temporal constraints.

Mixed-Integer and Stochastic Structures

For contact/motion planning, outer-level discrete schedules (e.g., number of contacts, terrain patch selections) are optimized using population-based stochastic search (e.g., Cross-Entropy Method), with the inner program a continuous OCP (e.g., trajectory, forces/tensions) solved by NLP methods (Malacarne et al., 29 Apr 2026). In this structure, adaptive sampling and elite-based update of the candidate distribution in the ULP efficiently explores combinatorial scenarios, while the LLP supplies continuous feasibility and cost metrics.

Implicit Dynamics and Physics Solvers as LLP

In complex robotic domains where system dynamics themselves require solution of contact, complementarity, or physics-based subproblems, the LLP at each time step is an optimization-based dynamics projection problem (Howell et al., 2021). The ULP (e.g., iLQR, DDP, shooting) solves for a sequence of control actions or trajectory states, while the LLP per step solves: $$\min_{\zeta \in \mathcal{Z}^{UL}} F\left(\zeta, x^*(\zeta)\right), \quad \mathrm{where}\quad x^*(\zeta) = \arg\min_{x \in \mathcal{X}^{LL}} f(\zeta, x) \ \text{s.t.} \ g_{LL}(\zeta,x) \leq 0$$3 Implicit differentiation of the KKT system for the interior-point solver provides smooth state/control Jacobians for use in outer-level optimization.

3. Hypergradient Computation: Analytical and Implicit Methods

Accuracy and efficiency of the outer-loop optimization hinge on computation of hypergradients (upper-level sensitivities). Several techniques are employed across TABL frameworks:

• Analytical KKT-based derivatives: As in (Chen et al., 2022), the exact solution of the equality-constrained QP admits closed-form gradients via solution of linear systems involving KKT multipliers.
• Sensitivity analysis for parametric QP/NLP: Application of classical results (e.g., Fiacco, Jittorntrum) yields expressions based on inner dual variables and explicit derivatives of cost/constraint matrices (Sun et al., 2018).
• Danskin’s theorem: For robust trajectory-optimization-with-policy (Kolaric et al., 2020), worst-case cost terms are differentiated using Danskin’s theorem through eigenvalue problems.
• Implicit Function Theorem: When LLP solutions are defined via KKT optimality, the Jacobian of the solution map is obtained by solving a linear system involving the KKT Jacobian, as in optimization-based dynamics solvers (Howell et al., 2021) or differentiable NLS pose solvers (Manoharan et al., 2024).
• Iterative approximate differentiation: In ML domains, iterative differentiation or unrolling (with or without truncation and initialization augmentation) is used to obtain approximate hypergradients through nonconvex and non-smooth LL trajectories (Liu et al., 2023, Liu et al., 2021).

4. Representative Applications

Trajectory-adaptive bi-level optimization has enabled significant advances across diverse control, planning, and learning domains:

• Minimum-time and minimum-snap UAV path planning: Dramatic order-of-magnitude speedup (from exponential to linear scaling in the number of safe sets/waypoints), with near-identical optimized costs (Chen et al., 2022, Sun et al., 2018).
• Redundant manipulator trajectory optimization: Bi-level QP-based formulations for joint trajectory and path speed under tight actuator constraints, with primal-dual upper-level iteration yielding sub-second QP solves and 10× speedup over active-set/IP methods (Fried et al., 2024, Fried et al., 4 Jun 2025).
• Autonomous vehicle navigation in dense traffic: Joint optimization of high-level behavioral setpoints (lane, speed) and low-level trajectory via GPU-accelerated batched convex optimization, with learned CVAE warm-starts for robust adaptation in dynamic scenes (Singh et al., 2022).
• Footstep/contact planning for legged robots: Mixed-integer CEM-based outer search coupled with continuous OCP inner solvers, enabling terrain-adaptive, energy-efficient multi-jump plans (Malacarne et al., 29 Apr 2026).
• Probabilistic adaptive trajectory control: Online learning of GP-based models (outer), with belief-space DDP for optimal action planning (inner) under model uncertainty, supporting efficient adaptation and real-time planning (Pan et al., 2016).
• Learning and meta-learning for hyperparameter or architecture search: Iterative trajectory modeling with trajectory-truncated hypergradient computations enables efficient, convergent bi-level solvers for nonconvex, high-dimensional problems (Liu et al., 2023, Wang et al., 2024).

5. Theoretical Properties and Convergence

• Convex LLPs: Many trajectory-adaptive schemes exploit problem reformulation to render the lower level a convex QP (or even analytic maximum-of-quadratics), ensuring existence, uniqueness, and efficient solution. Convergence guarantees at the upper level follow from subdifferential calculus and epi-convergence under mild regularity (Fried et al., 2024, Fried et al., 4 Jun 2025).
• Nonconvex/Non-smooth LLPs: For bilevels with nonconvex or non-smooth LLs, the theory of iterative hypergradient computation with initialization auxiliary and pessimistic trajectory truncation ensures global convergence to stationary points, even without lower-level convexity (Liu et al., 2023, Liu et al., 2021).
• Sensitivity accuracy: Analytical gradients based on KKT and implicit differentiation have been empirically shown to outperform finite-difference approaches by an order of magnitude in accuracy and speed, with better scaling and stability in high-dimensional problems (Sun et al., 2018, Chen et al., 2022).
• Scalability: Complexity per iteration is typically dominated by single or batched QPs (cubic in variable count, but linear in the number of segments/safe sets/waypoints), while upper-level gradient steps scale linearly in the number of constraints (Chen et al., 2022, Singh et al., 2022).

6. Practical Impact and Limitations

• Real-time deployment: Several frameworks achieve sub-20 ms update rates, with on-board operation for UAVs and mobile robots in dynamic environments, supported by efficient sparse/parallelizable QP routines (Sun et al., 2018, Singh et al., 2022, Manoharan et al., 2024).
• Quantitative gains: In UAV planning, order-of-magnitude speedups in computational time with maintained or minimally higher final cost have been repeatedly validated (e.g., 141 ms vs. 361 ms for two-waypoint test, 363 ms vs. 4.3 s for three-waypoint dynamic sets (Chen et al., 2022)).
• Robustness to complexity: The separation of trajectory parameters and physical or safety constraints—together with efficient gradient computation—enables robust handling of terrain, contact, collision, or dynamics constraints beyond what is possible for monolithic NLP transcriptions.
• Model and solver assumptions: Analytical sensitivity and the guarantees of convex LLPs require regularity (e.g., strong convexity, strict feasibility, constraint qualifications). For nonconvex LLPs, convergence to global optima cannot be ensured without further structure (Liu et al., 2023, Liu et al., 2021).
• Learning-driven architectures: For nonconvex, high-dimensional ML domains, AIT and related frameworks extend the reach of TABL to neural architecture search, meta-learning, robust learning, and adversarial games, subject to computational overhead from unrolling and increased memory usage (Liu et al., 2023).

7. Summary Table: Paradigms of Trajectory-Adaptive Bi-Level Optimization

Approach / Domain	LLP Structure	ULP Variable(s)	Gradient Method
UAV min-snap path (Chen et al., 2022)	Equality QP	Waypoints, Timing	KKT/OptNet analytic
Redundant manipulator (Fried et al., 2024)	Convex QP / Closed	Trajectory coefficients	Primal-dual, subgradient
Contact planning (legged) (Malacarne et al., 29 Apr 2026)	NLP (OCP)	Contacts, patches	CEM, adjoint/NLP
Learned dynamics (Pan et al., 2016)	GP/Belief-space	Model parameters	Analytic, DDP
ML/Meta-learning (Liu et al., 2023, Liu et al., 2021)	Iterative, nonconvex	Hyperparams, LL init	Unroll, truncated backprop
Automated driving (Singh et al., 2022)	Convex (SCP) batch	Behavioral setpoints	CVAE, CEM, GPU batch
Trajectory via Koopman (Abou-Taleb et al., 2024)	Linear QP in lift	Boundary, T	Finite-diff or adjoint

Each line abstracts a distinct instantiation of the trajectory-adaptive bilevel strategy, tailored for computational tractability, constraint handling, and sensitivity accuracy within the problem’s domain.

---

In modern optimal control, planning, and learning, trajectory-adaptive bi-level optimization offers a unifying formalism for decomposing, solving, and adapting complex trajectory-centric problems with constraints, uncertainty, and high dimensionality, using both classical and learning-driven tools (Chen et al., 2022, Sun et al., 2018, Fried et al., 2024, Fried et al., 4 Jun 2025, Malacarne et al., 29 Apr 2026, Liu et al., 2023, Singh et al., 2022, Howell et al., 2021).