HyperMPC: Predictive Control with HyperPM

• HyperMPC is a model predictive control framework that uses the HyperPM neural architecture to predict a trajectory of time-varying parameters over the control horizon.
• It projects unmodeled dynamics onto a lightweight analytical model, balancing computational efficiency with enhanced anticipatory control.
• Empirical results in simulated pendulum, drone payload, and autonomous racing demonstrate its superior performance compared to static and residual models.

Searching arXiv for the primary HyperMPC paper and related mentions to ground the article in current literature. HyperMPC is a model predictive control framework built around the Hyper Prediction Model (HyperPM), a neural architecture that predicts a trajectory of time-varying parameters for a base dynamics model across the MPC prediction horizon. Instead of replacing the nominal model with a purely residual or black-box dynamics predictor, it projects unmodeled dynamics onto a time-dependent parameterization of an otherwise lightweight analytical or hybrid model. In the reported formulation, this preserves the computational efficiency and robustness of the base model while enabling anticipation of previously unmodeled phenomena; when integrated into MPC, the method consistently outperforms existing state-of-the-art techniques on a simulated pendulum with backlash, a drone with a rope-suspended payload, and real F1TENTH autonomous racing (Węgrzynowski et al., 8 Aug 2025).

1. Origin and problem setting

HyperMPC was introduced to address a specific limitation of gradient-based model predictive control: existing dynamics models used in such controllers are constrained by computational complexity and state representation. The motivating cases are systems in which relevant dynamics are unmodeled, unobserved, or only partially observed. Examples given for this setting include payload swing in a drone with a rope-suspended load and tire-related effects in racing; in such systems, longer-horizon predictive accuracy degrades when the controller relies on static parameters or static model adaptations (Węgrzynowski et al., 8 Aug 2025).

The central methodological departure is from constant-parameter adaptation. Prior approaches described in the source material either use fully analytical models with constant parameters, hybrid or data-driven models with static parameters over the optimization horizon, or hyper-network-style adaptations that predict only static parameters across the horizon. HyperMPC instead forecasts how model parameters should evolve over the horizon as a function of both recent observations and anticipated future controls. This suggests that the framework is aimed not merely at online identification of the current regime, but at preemptive compensation for near-future dynamical effects.

This design choice is consequential because MPC evaluates candidate future control sequences against an internal model. If that internal model can be adjusted in a horizon-aware manner before optimization, then predictive inaccuracies induced by hidden or unmodeled dynamics can be shifted from the optimizer into a learned parameter trajectory, while the optimizer itself remains structurally unchanged.

2. Mathematical formulation

The base parameterized dynamics model is written as

$x = f_\theta(x, u),$

where $\theta$ is the parameter vector. In the source formulation, traditional system identification is cast as

$$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$

HyperPM replaces the constant-$\theta$ assumption with a horizon-level parameter trajectory. The predicted parameter sequence is

$$\theta[t_c, t_c + t_p] = \text{HyperPM}\big( x[t_c-t_h, t_c], \; u[t_c-t_h, t_c], \; \hat{u}[t_c, t_c + t_p] \big).$$

The inputs are past state history, past control history, and anticipated future controls. The output is a sequence of model parameters over the prediction horizon (Węgrzynowski et al., 8 Aug 2025).

The resulting state rollout uses time-varying parameters: $\hat{x}_{t+1} = f_{\theta_t}(\hat{x}_t, u_t).$ In this construction, unmodeled dynamics are not represented as an additive residual term on the state transition, but as a sequence of parameter deviations injected into the base model at each horizon step. The paper characterizes this as projecting the influence of unmodeled dynamics onto time-varying parameters of an efficient analytical or hybrid model.

A plausible implication is that HyperMPC preserves the solver-facing structure of the nominal control problem more directly than residual-model alternatives. The paper explicitly frames this as a way to retain computational efficiency and robustness while increasing anticipatory capacity.

3. HyperPM architecture and training

HyperPM comprises four stated components. The first is a history encoder that recovers a latent state from past $x$ and $u$; a GRU was found empirically superior to LSTM and MLP alternatives. The second is a future-controls preprocessor that represents anticipated controls using B-spline control points. This is described as a dimensionality-reduction and robustness device, and as a way to focus on control trends rather than sampling-level details. During training, random Gaussian noise $\bigl(N(1, \sigma_{\text{robust}})\bigr)$ is injected into the B-spline points to improve robustness to mismatch between planned and actual controls (Węgrzynowski et al., 8 Aug 2025).

The third component is a causal parameter predictor, implemented as a two-layer masked MLP. It receives the latent state and processed future controls and outputs parameter deviations over time. The masking enforces causality, so that future actions cannot influence past parameter estimates. The fourth component is parameter interpolation: B-spline interpolation produces smooth parameter trajectories $\theta[t_c, t_c + t_p]$, with scaling $\theta$0 and offset $\theta$1 hyperparameters controlling their range and reference values.

Regularization is applied through an $\theta$2 penalty on parameter changes, weighted by $\theta$3. The training loss is reported as

$\theta$4

Optimization uses backpropagation through time. During training, true future controls from data are provided as inputs, and long-horizon errors are directly minimized. During deployment, the model receives predicted future controls from the previous MPC iteration or from a planner. The injected control noise is explicitly motivated as a countermeasure to this train/test mismatch (Węgrzynowski et al., 8 Aug 2025).

This training protocol is significant because the model is optimized for the same long-horizon rollout behavior that matters to MPC, rather than only for single-step prediction accuracy.

4. Integration into the MPC loop

HyperMPC embeds HyperPM into a standard optimization-based MPC cycle. At each control step, the most recent state and control history together with previously planned future controls are passed to HyperPM, which returns a parameter trajectory over the horizon. The nominal model is then augmented so that each step in the horizon uses the corresponding predicted parameter value. The MPC problem is solved as usual, including implementations via acados or other solvers, but with a time-varying internal model. Only the first control is executed before the cycle repeats (Węgrzynowski et al., 8 Aug 2025).

The reported computational consequence is narrowly defined: only a single forward pass through the HyperPM network occurs before the MPC solve, and the complexity of the optimization itself is unchanged compared with standard MPC using the same analytical model. The source also reports only a marginal increase in parameter-inference time, for example less than $\theta$5 ms, with solver time unchanged compared to static models.

This point distinguishes HyperMPC from approaches that increase the burden inside the optimizer itself. The framework is presented as changing the model instance that the optimizer sees, not the optimizer’s mathematical structure. As a result, the claimed advantage is not only improved prediction fidelity, but also preservation of the real-time characteristics associated with lightweight analytical MPC models.

5. Empirical evaluation

The reported evaluation covers three domains chosen for substantial unmodeled or hidden effects: a simulated pendulum with backlash, a simulated drone with a rope-suspended payload, and real F1TENTH autonomous racing. Baselines include constant-parameter models trained on single or long sequences, HyperDynamics variants with horizon-static parameters, residual models, and oracle adaptation where true hidden effects are available from the simulator and treated as constant over the horizon (Węgrzynowski et al., 8 Aug 2025).

Domain	Prediction result	Control result
Pendulum with backlash	$\theta$6; HyperPM error $\theta$7	Successful swing-up and stabilization in all episodes; cost $\theta$8 vs $\theta$9 for best baseline; zero solver failures
Drone with rope-suspended payload	$$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$0 (nominal); $$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$1 (residual); HyperPM error $$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$2	Lower trajectory tracking cost than HD and oracle adaptation; benefit increases with more significant unmodeled dynamics
Real F1TENTH racing	$$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$3 vs const; $$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$4 vs HD; HyperPM error $$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$5	Lowest racing cost and safety penalties; outperforms residual models known for superior open-loop prediction but poor generalization

Several ablations are reported. Conditioning on future controls is described as critical: omitting it degrades long-horizon prediction, by up to $$\min_\theta \; \int_{t_c}^{t_c + t_p} \| x(t) - \hat{x}(t) \|^2 \, dt \quad \text{s.t.} \quad \hat{x} = f_\theta(\hat{x}, u).$$6 for F1TENTH. Time-varying parameterization is also reported as essential. Even oracle adaptation with perfect knowledge of the current hidden state is outperformed, because HyperPM can forecast how unmodeled effects evolve over the horizon. The source further states that residual models may show superior open-loop prediction, but exhibit poor generalization in the control setting (Węgrzynowski et al., 8 Aug 2025).

Taken together, these results position HyperMPC as a method whose benefit appears most strongly when hidden-state influence grows over the horizon, rather than when the task can be handled by a static local correction.

6. Interpretation, scope, and nomenclature

A central interpretive claim in the source material is that HyperMPC remains lightweight and interpretable. Its parameter trajectories are described as physically meaningful; in F1TENTH, parameter evolution is said to resemble true physical effects such as load transfer. The framework therefore occupies an intermediate position between fixed analytical modeling and unconstrained residual learning: it uses a neural network, but the network predicts how to deform a structured dynamics model rather than directly replacing it (Węgrzynowski et al., 8 Aug 2025).

This directly addresses two common misunderstandings. First, HyperMPC is not presented as a residual-MPC architecture. The learned object is a time-varying parameter trajectory, not an additive state-transition correction. Second, it is not merely an online constant-parameter adaptation method. The defining feature is horizon-level prediction of parameter evolution conditioned on planned controls. This is why the paper emphasizes anticipation of nonlinear, irreversible, or delayed effects rather than only adaptation to the current hidden state.

The term itself is not globally unique across arXiv-adjacent usage. In the provided materials, the label “HyperMPC” also appears in summaries of unrelated secure multiparty computation work, including large-scale privacy-preserving biometric verification and blockchain-oriented secure computation systems (Bloemen et al., 2024, Zhang et al., 2018). This suggests a nomenclatural overlap rather than a shared technical lineage. Within control, however, HyperMPC denotes the integration of HyperPM into model predictive control as introduced in “Beyond Constant Parameters: Hyper Prediction Models and HyperMPC” (Węgrzynowski et al., 8 Aug 2025).

For control research, the framework’s main significance lies in its claim that long-horizon anticipatory modeling can be improved without altering the computational complexity of the underlying MPC solve. That places HyperMPC in a specific design space: analytical-model MPC augmented by learned, time-varying parameter trajectories driven by history and future control plans, with empirical gains demonstrated on systems whose dominant errors arise from hidden or partially modeled dynamics.