BAN-MPC: Barrier-integrated Adaptive Neural MPC

• BAN-MPC is a control framework that integrates Control Barrier Functions with a short-horizon MPC, augmented by a learned neural terminal cost and sensitivity network to ensure safety under varying conditions.
• It shifts long-horizon dynamics reasoning into offline-learned value functions, achieving over 200× speedup on embedded platforms while maintaining strict safety via constraint handling.
• Experiments demonstrate collision-free navigation and control errors below 5% under 15% parameter variations, proving its efficacy and adaptability in complex nonlinear systems.

BAN-MPC most commonly denotes Barrier-integrated Adaptive Neural Model Predictive Control, a control framework that couples strict safety via Control Barrier Functions (CBFs) with short-horizon Model Predictive Control (MPC) augmented by an offline-learned neural value function and a second neural network that learns the sensitivity of that value function to system parameters (Wang et al., 8 Sep 2025). It is intended for regimes in which conventional MPC can enforce safety through constraints but its real-time execution exceeds embedded compute budgets. In the reported hardware-in-the-loop (HIL) experiments on Jetson Nano, BAN-MPC solves 200 times faster than traditional MPC, enables collision-free navigation, and maintains control error below $5\%$ under model parameter variations within $15\%$ (Wang et al., 8 Sep 2025).

1. Design objective and problem setting

BAN-MPC is proposed for discrete-time nonlinear systems with parameters $p$ (or $\theta$ in the paper), state $x \in \mathbb{R}^{n_x}$, and input $u \in \mathbb{R}^{n_u}$, governed by

$x_{k+1} = f(x_k, u_k, p).$

The motivating difficulty is threefold. First, conventional CBF-MPC with long horizons and nonlinear dynamics can be prohibitively slow on embedded platforms such as Jetson Nano. Second, purely neural approximations, described in the source as AMPC, accelerate computation but sacrifice strict safety guarantees. Third, short-horizon MPC reduces runtime but is myopic and loses long-horizon performance (Wang et al., 8 Sep 2025).

The central design choice is to retain optimization-based constraint handling while shifting long-horizon reasoning into learned terminal structure. BAN-MPC therefore uses a Short-horizon MPC with horizon $M \ll N$, embeds an offline-learned neural value function as terminal cost, and adapts that terminal value online when model parameters change. The result is a controller characterized in the source as fast, safe, and adaptive (Wang et al., 8 Sep 2025).

A common misconception is to view BAN-MPC as a direct policy approximation. The formulation does not replace the optimization with a policy network. Instead, it preserves the MPC problem itself, including the CBF constraints, and learns only the terminal cost and its parameter sensitivity. This distinction is central to why the method preserves forward-invariance-based safety guarantees under nominal dynamics while still reducing online complexity (Wang et al., 8 Sep 2025).

2. Optimization formulation

The long-horizon expert problem used for supervision is a classical nonlinear CBF-MPC. Its stage cost is

$\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$

with state and input constraints $x_k \in \mathcal{X}_k$, $15\%$0, measured or estimated initial condition $15\%$1, and discrete-time CBF constraints

$15\%$2

where typically $15\%$3 and $15\%$4. The expert optimization is

$15\%$5

subject to the dynamics, state/input constraints, initial condition, and the CBF constraints above. In the baseline CBF-MPC used to generate expert labels, no learned terminal cost is appended, because the horizon is long enough to approximate the cost-to-go directly (Wang et al., 8 Sep 2025).

BAN-MPC replaces that long-horizon problem with a short-horizon problem of horizon $15\%$6, for example $15\%$7, and appends an adapted terminal value: $15\%$8 The resulting optimization is

$15\%$9

subject to

$p$0

$p$1

$p$2

$p$3

The intended effect is to embed learned long-horizon cost-to-go information into a compact online problem so that BAN-MPC recovers long-horizon behavior at a fraction of the computational cost (Wang et al., 8 Sep 2025).

The complete formulation in the source makes the adaptation explicit: $p$4 where $p$5 is the learned sensitivity map. This identifies BAN-MPC as a hybrid of nonlinear MPC, neural terminal-value approximation, and first-order parametric sensitivity correction (Wang et al., 8 Sep 2025).

3. Safety mechanism: Control Barrier Functions and obstacle composition

A defining feature of BAN-MPC is the replacement of Euclidean distance penalties or ad hoc soft constraints by Control Barrier Functions used as hard constraints. The safe set is

$p$6

and the discrete-time CBF constraint is

$p$7

In continuous time the standard form is

$p$8

Within MPC, enforcing the discrete-time constraint at each predicted step yields collision avoidance under nominal dynamics (Wang et al., 8 Sep 2025).

For multi-obstacle environments, the framework uses a smooth soft-min over obstacle-wise barrier functions: $p$9 The source states that this ensures twice continuous differentiability, which is needed for sensitivity/KKT assumptions and efficient nonlinear programming. For circular obstacles, unicycle radius $\theta$0, obstacle center $\theta$1, obstacle radius $\theta$2, and safety margin $\theta$3, the obstacle-wise barrier is

$\theta$4

so that $\theta$5 means the robot is outside the inflated obstacle (Wang et al., 8 Sep 2025).

The forward-invariance guarantee is the standard discrete-time CBF lemma: if $\theta$6 and the input sequence enforces

$\theta$7

for all $\theta$8, then $\theta$9 is forward invariant. In receding-horizon form, safety is preserved provided feasibility holds. This also clarifies a second common misconception: the safety guarantee is not derived from the learned neural components, but from the barrier constraints retained inside the optimization (Wang et al., 8 Sep 2025).

The same construction extends to moving obstacles through time-varying barrier functions $x \in \mathbb{R}^{n_x}$0 built from predicted obstacle centers $x \in \mathbb{R}^{n_x}$1, with the constraint

$x \in \mathbb{R}^{n_x}$2

This suggests that BAN-MPC is structurally compatible with dynamic-obstacle settings as long as obstacle prediction is reliable (Wang et al., 8 Sep 2025).

4. Learned terminal value and sensitivity-based adaptation

The first learned component, $x \in \mathbb{R}^{n_x}$3, approximates the optimal cost-to-go $x \in \mathbb{R}^{n_x}$4 of the long-horizon CBF-MPC problem. It is trained offline using expert data generated by solving the long-horizon CBF-MPC over diverse initial conditions and obstacle configurations. In the reported experiments, the expert horizon is $x \in \mathbb{R}^{n_x}$5, and the dataset contains tuples $x \in \mathbb{R}^{n_x}$6. The training loss is the mean-squared error

$x \in \mathbb{R}^{n_x}$7

Its role is purely terminal: a single forward pass replaces explicit long-horizon rollout inside the online optimization (Wang et al., 8 Sep 2025).

The source emphasizes Value Function Dataset Aggregation (VF-DAGGER) to mitigate distribution shift. Rather than performing one-shot behavioral cloning on the expert state distribution, the method iteratively embeds the current value approximation $x \in \mathbb{R}^{n_x}$8 as terminal cost in short-horizon MPC to form policy $x \in \mathbb{R}^{n_x}$9, rolls out $u \in \mathbb{R}^{n_u}$0 for $u \in \mathbb{R}^{n_u}$1 steps, labels the visited states with expert MPC, aggregates those labels into the dataset, and retrains. An optional mixing step

$u \in \mathbb{R}^{n_u}$2

combines the initial value with the retrained estimate. The appendix guarantees reported for VF-DAGGER state that this iterative aggregation avoids the $u \in \mathbb{R}^{n_u}$3 error growth typical of naive behavioral cloning and yields a sublinear performance gap relative to expert CBF-MPC (Wang et al., 8 Sep 2025).

The second learned component is the sensitivity network

$u \in \mathbb{R}^{n_u}$4

where $u \in \mathbb{R}^{n_u}$5 denotes dynamics parameters such as mass, friction, or arm lengths. BAN-MPC then uses the first-order adaptation law

$u \in \mathbb{R}^{n_u}$6

The source ties this approximation to implicit-function-theorem and KKT sensitivity analysis for the parametric nonlinear program

$u \in \mathbb{R}^{n_u}$7

with Lagrangian

$u \in \mathbb{R}^{n_u}$8

Under LICQ, SOSC, and strict complementarity, and under small parameter perturbations that keep the active set unchanged, the Implicit Function Theorem yields first-order sensitivity of the KKT point and of the optimal value (Wang et al., 8 Sep 2025).

This adaptation is specifically designed to avoid retraining when parameters vary. At runtime BAN-MPC evaluates $u \in \mathbb{R}^{n_u}$9 and $x_{k+1} = f(x_k, u_k, p).$0, computes $x_{k+1} = f(x_k, u_k, p).$1 in $x_{k+1} = f(x_k, u_k, p).$2, and thereby avoids both retraining and the combinatorial offline dataset expansion that would arise from gridding over parameter values. The limitation stated in the source is equally important: if parameter variations exceed the trained bounds or the active constraint set changes, the linear sensitivity approximation may become inaccurate (Wang et al., 8 Sep 2025).

5. Theoretical properties, workflow, and empirical performance

The theoretical claims in the source are local and assumption-driven. The assumptions include continuity of the dynamics model $x_{k+1} = f(x_k, u_k, p).$3, compact and regular constraints $x_{k+1} = f(x_k, u_k, p).$4, twice continuously differentiable $x_{k+1} = f(x_k, u_k, p).$5, sufficiently small bounded parameter variations so that active constraint sets remain unchanged, and bounded learned value and sensitivity errors with high probability (Wang et al., 8 Sep 2025).

On that basis, BAN-MPC inherits forward invariance from the CBF constraints and admits a standard recursive-feasibility sketch: if a feasible plan over $x_{k+1} = f(x_k, u_k, p).$6 steps exists and the first control $x_{k+1} = f(x_k, u_k, p).$7 is applied, then safety is maintained and the shifted problem remains feasible under nominal assumptions. The appendix also includes a Probabilistic Practical Exponential Stability theorem. Under

$x_{k+1} = f(x_k, u_k, p).$8

$x_{k+1} = f(x_k, u_k, p).$9

bounded approximation errors for $M \ll N$0 and $M \ll N$1, bounded parameter variations, and a Lipschitz-like relation between BAN-MPC and MPC value differences, one obtains

$M \ll N$2

with probability at least $M \ll N$3. The radius $M \ll N$4 scales with approximation and parameter mismatch terms. This is a practical rather than exact stability result, and it explicitly depends on learned approximation quality and moderate parametric shifts (Wang et al., 8 Sep 2025).

The deployment workflow is correspondingly structured. Offline, one generates expert data from long-horizon CBF-MPC, trains $M \ll N$5, runs VF-DAGGER, builds a sensitivity dataset $M \ll N$6 using automatic differentiation, finite differences, or analytic/KKT-based sensitivity, and trains $M \ll N$7. Online, one senses or estimates current $M \ll N$8 and $M \ll N$9, computes

$\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$0

solves the short-horizon MPC with CBF constraints, applies $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$1, shifts the horizon, and repeats. The computational rationale is direct: long-horizon CBF-MPC has $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$2 stages with nonlinear and barrier constraints at each stage, whereas BAN-MPC uses $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$3 stages with $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$4 plus two neural forward passes and a vector dot product (Wang et al., 8 Sep 2025).

The reported HIL experiments use a host PC and Jetson [email protected] with Ubuntu 18.04, Torch 1.10, CasADi 3.5, and IPOPT. The neural networks have three hidden layers with 32 tanh units each. Data generation uses $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$5 samples for a unicycle and $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$6 samples for a quadrotor, generated offline on a 128-core AMD Ryzen 9 5995WX workstation. Baselines are CBF-MPC ($\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$7), Short-horizon MPC ($\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$8), AMPC, Neural MPC, and BAN-MPC (Wang et al., 8 Sep 2025).

For the unicycle, with five circular obstacles, robot radius $\ell(x_k, u_k) = x_k^\top Q x_k + u_k^\top R u_k,$9, safety margin $x_k \in \mathcal{X}_k$0 m, and $x_k \in \mathcal{X}_k$1, BAN-MPC tracks the CBF-MPC trajectory closely, Short-horizon MPC fails, and AMPC shows unsafe behavior. On Jetson Nano, the reported metrics are domain safety $x_k \in \mathcal{X}_k$2, average suboptimality $x_k \in \mathcal{X}_k$3 versus CBF-MPC, and average compute time $x_k \in \mathcal{X}_k$4 s versus $x_k \in \mathcal{X}_k$5 s for CBF-MPC. For the quadrotor, BAN-MPC remains close to CBF-MPC while AMPC collides and Short-horizon MPC deviates heavily; compute times are $x_k \in \mathcal{X}_k$6 s for CBF-MPC, $x_k \in \mathcal{X}_k$7 s for AMPC, and $x_k \in \mathcal{X}_k$8 s for BAN-MPC, corresponding to $x_k \in \mathcal{X}_k$9 speedup versus MPC, with resource utilization $15\%$00. Under parameter deviations $15\%$01 from nominal, BAN-MPC preserves safety and maintains control error below $15\%$02 (Wang et al., 8 Sep 2025).

These results support a narrow but important conclusion. BAN-MPC is not presented as a universally robust surrogate for nonlinear MPC under arbitrary uncertainty; rather, it is a short-horizon, CBF-constrained, value-augmented controller whose empirical gains are strongest on embedded hardware and whose adaptation mechanism is designed for moderate parameter changes without retraining (Wang et al., 8 Sep 2025).

6. Distinct acronym usages and adjacent meanings

The acronym BAN-MPC is not semantically uniform across the broader literature represented in the supplied sources. In one adjacent use, it is interpreted as a best-of-both-worlds MPC framework rather than as Barrier-integrated Adaptive Neural Model Predictive Control. The relevant paper, "Perfectly Secure Synchronous MPC with Asynchronous Fallback Guarantees Against General Adversaries" (Appan et al., 2022), explicitly states that it does not use the term “BAN-MPC”; instead, the supplied interpretation treats BAN-MPC as “best-of-both-worlds MPC with BA and VSS components that yield synchronous behavior when the network is synchronous and safe asynchronous fallback when the network is asynchronous.” Under that interpretation, the protocol is a single, perfectly secure MPC framework secure against $15\%$03 adversary structures in synchronous networks and $15\%$04 adversary structures in asynchronous networks, without parties knowing the network type in advance (Appan et al., 2022).

A second adjacent use connects BAN-MPC to banded ADMM for standard MPC. The paper "Efficient implementation of MPC for tracking using ADMM by decoupling its semi-banded structure" (Gracia et al., 2024) presents an ADMM-based solver for MPC for tracking and explains it as a direct generalization of BAN-MPC in the sense of banded ADMM for regulation MPC. In that discussion, regulation MPC yields strictly banded KKT systems, whereas tracking MPC introduces artificial steady-state variables that produce low-rank off-band couplings. The paper restores banded subproblems using a Woodbury-based decomposition so that the resulting complexity remains $15\%$05 per ADMM iteration for fixed $15\%$06 (Gracia et al., 2024).

These usages are conceptually unrelated. The first concerns information-theoretic secure multi-party computation under mixed synchronous/asynchronous network assumptions. The second concerns numerical linear algebra and operator splitting for optimization-based control. The third, and the primary meaning in contemporary control usage, denotes the barrier-integrated adaptive neural controller introduced in "Safety Meets Speed: Accelerated Neural MPC with Safety Guarantees and No Retraining" (Wang et al., 8 Sep 2025). A plausible implication is that the acronym should be expanded explicitly in technical writing, because “BAN-MPC” alone does not uniquely identify a single research program across neighboring literatures.