Sampling-Based Predictive Control

Updated 29 June 2026

Sampling-based Predictive Control is a family of model predictive control methods that uses forward sampling of candidate control trajectories to approximate optimal solutions in nonlinear and high-dimensional systems.
It employs various techniques such as random shooting, CEM, MPPI, and trust-region updates to iteratively refine the control distribution without relying on gradient-based methods.
Its practical applications span robotics, dexterous manipulation, and contact-rich tasks, offering robust performance under uncertainty and dynamic constraints.

Sampling-based Predictive Control (SBPC) denotes a family of model predictive control (MPC) methods that use forward sampling of control trajectories in a dynamical system to approximate the optimal solution to finite-horizon optimal control problems, particularly in settings with nonlinear, non-smooth dynamics or high-dimensional state and action spaces. Instead of relying on gradient-based optimization or explicit system linearization, SBPC exploits stochastic or deterministic sampling of candidate input sequences, simulated rollouts via black-box or physics-based models, and cost-weighted selection or averaging to update the control distribution in a receding-horizon fashion. These methods span a spectrum from basic random shooting or predictive sampling to advanced path integral control, cross-entropy method (CEM) MPC, and newer trust-region or learning-augmented variants. SBPC is central to contemporary real-time robotics, high-DOF manipulation, contact-rich tasks, and stochastic MPC under uncertainty.

1. Core Problem Formulation and Theoretical Underpinnings

SBPC approaches the discrete-time, finite-horizon optimal control problem for a possibly nonlinear or black-box dynamic system: $x_{k+1} = f(x_k, u_k), \quad x_k \in \mathbb{R}^n,\quad u_k\in\mathbb{R}^m$ with a cost functional

$J(U) = \mathbb{E}\left[\sum_{k=0}^{N-1} \ell(x_k, u_k) + \Phi(x_N)\right]$

where the expectation is optionally over model, parameter, or process uncertainty, and the decision variable is the control sequence $U = \{u_0, u_1, ..., u_{N-1}\}$ .

Unlike classic gradient-descent MPC, SBPC does not require differentiability of $f$ or $\ell$ . Instead, it maintains a parameterized proposal distribution (often Gaussian) over $U$ , samples multiple candidate sequences, evaluates their costs via either physics simulation or model rollouts, and uses these costs to steer the proposal distribution toward lower-cost regions. The process is iterated either once per control step, or via a fixed number of inner iterations before receding the horizon.

Key theoretical machinery includes:

Monte Carlo importance sampling: SBPC estimates the expected cost or optimality statistics via parallel simulation of sampled trajectories.
Path-integral (PI) and information-theoretic control: In variants such as MPPI, the path integral of cost-to-go under the sampling distribution yields a closed-form expression for the cost-weighted mean trajectory, effectively performing a zeroth-order natural gradient step (Pezzato et al., 2023, Schramm et al., 24 Nov 2025).
Receding-horizon update: Only the first input in the best or mean sequence is applied, after which the distribution is shifted forward for the next iteration.

2. Algorithmic Taxonomy and Representative Methods

SBPC encompasses several algorithmic templates, varying in their sampling/distribution update rules and optimization sophistication.

Predictive Sampling: The simplest approach, as implemented in DeepMind’s MuJoCo MPC framework, perturbs the nominal control sequence or compressed spline knot vector with i.i.d. noise, simulates each candidate, and selects the control achieving lowest cost (Howell et al., 2022). This approach is fundamentally derivative-free, myopic, and embarrassingly parallel, but may require significant sampling for high-dimensional problems.

Cross-Entropy Method (CEM-MPC): CEM iteratively fits a Gaussian (or mixture) to the elite set of lowest-cost trajectories sampled in each iteration, updating the mean and covariance toward low-cost regions. This update can be performed with hard “elite” sets or soft exponential reweighting (Li et al., 20 Jun 2025, Walker et al., 7 Jan 2026).

Model Predictive Path Integral (MPPI): MPPI interprets the control optimization as the minimization of a free-energy functional; candidate control sequences are sampled around a nominal plan and reweighted according to an exponential of negative cost divided by a temperature parameter. The update is a cost-weighted average, giving a softmin-shift toward the minima in the sampled set (Pezzato et al., 2023, Schramm et al., 24 Nov 2025, Sacks et al., 2022). Variants such as Biased-MPPI allow arbitrary proposals and the fusion of ancillary controllers (Trevisan et al., 2024).

Trust-Region Sampling-based MPC: This variant constrains proposal distribution updates via explicit Kullback–Leibler divergence bounds, optionally with entropy lower bounds. It removes the need for hand-tuned temperature schedules, solving a dual concave program to determine optimal update parameters, and leverages deterministic low-discrepancy sampling for increased efficiency (Walker et al., 8 May 2026, Walker et al., 7 Jan 2026).

Learning-based and Latent-space Sampling: Recent work embeds all sampling and warm-start logic in a learned latent space, often via normalizing-flow parameterizations, enabling rapid adaptation and better sampling in complex, multimodal cost landscapes (Sacks et al., 2022, Sacks et al., 2022).

3. Sampling Strategies: Structured, Adaptive, and Efficient

The dimension and structure of the sampled control parameterization critically determine sample efficiency, control smoothness, and convergence:

Unstructured (i.i.d.) Sampling: Perturbations are injected at every control timestep; simple but induces high variance and nonsmooth controls (Howell et al., 2022, Li et al., 20 Jun 2025).
Spline-based Sampling: Noise is injected only on compressed spline knots (e.g., cubic/Hermite/Bézier), then interpolated over the control horizon. This greatly reduces the search dimension and enforces controllable smoothness, which is essential for contact-rich robotics (Tao et al., 4 Jan 2026, Schramm et al., 24 Nov 2025). Cubic-spline parameterization with K knots (K ≪ N) was repeatedly found optimal in both performance and smoothness metrics for legged locomotion (Tao et al., 4 Jan 2026).
Colored and Low-pass Noise: Band-limited or low-pass filtered perturbations further constrain the frequency content of sampled controls, reducing chattering and actuator wear as in LP-MPPI (Kicki, 13 Mar 2025).
Deterministic Low-Discrepancy Sampling: Replaces random sampling with deterministic, quasi-Monte Carlo sample sets (e.g., LCD), reducing the variance of cost estimates and improving performance in the low-sample regime (Walker et al., 7 Jan 2026, Walker et al., 8 May 2026).

Empirically, structure in the sampling process – such as basis reduction and smoothness priors – is crucial for high-dimensional, contact-dense, or real-time applications.

4. Handling Uncertainty, Constraints, and Objective Adaptation

SBPC can natively handle model, parameter, and disturbance uncertainty by incorporating stochastic forward sampling (Monte Carlo) at two levels:

Parameter Uncertainty: For example, in robotic laser surgery, sampling-based MPC propagates model parameter uncertainties in absorption or threshold via ensemble rollouts, penalizing cost variance or explicitly computing expected penalized cost (Wang et al., 2024).
Chance Constraints and Safety: Hard constraints (e.g., obstacle avoidance, critical tissue preservation) are enforced by rejecting any rollout that violates state or input constraints. Soft constraints (overcuts, slight constraint violations) are penalized in the cost function with large but finite weights, maintaining exploratory behavior while discouraging unsafe solutions (Wang et al., 2024, Howell et al., 2022).
Objective Adaptation: SBPC facilitates rapid tuning of cost function weights via integration with automated systems, as demonstrated by the closed-loop adaptation using a visual LLM to optimize weights for task objectives in physical robotic hand manipulation (Hess et al., 2024).
Scenario and Chance-Constrained SBPC: In linear stochastic MPC, offline uncertainty sampling and corresponding constraint tightening enables probabilistic guarantees of constraint satisfaction and asymptotic stability (Lorenzen et al., 2016). Data-driven variants employ random disturbance/data parameterizations consistent with observed system responses, with recursive feasibility assured through robust first-step constraints and inner deterministic approximations of chance constraints (Teutsch et al., 2024, Teutsch et al., 2024).

5. Numerical Performance, Applications, and Empirical Insights

SBPC has been validated across a spectrum of high-dimensional, nonlinear, and contact-rich domains:

Robotic Laser Surgery: Random branching with cost-weighted expansion and rigorous constraint rejection/planning enables safe, precise, and uncertainty-aware volumetric ablation, automatically optimizing cut order, angles, powers, and exposure (Wang et al., 2024).
Dexterous Manipulation and Hands: Predictive Sampling combined with physics simulation and rapid cost-adaptation strategies delivered successful in-hand manipulation and catching on a biomimetic tendon-driven hand, with real-time iteration and sample counts as low as 10–25 per step (Hess et al., 2024).
Legged Locomotion and Humanoids: Real-time MPPI with structured dual-space spline parameterization enabled the emergence of diverse, reference-free behaviors including trotting, galloping, backflips, and dynamic handstands, at 50 Hz and with order-of-magnitude lower sample counts than prior GPU-based methods (Schramm et al., 24 Nov 2025). Spline-based sampling universally improved convergence, stability, and sample efficiency in simulation studies (Tao et al., 4 Jan 2026).
Whole-Body Manipulation, Navigation, and Racing: GPU-accelerated MPPI (IsaacGym) achieved state-of-the-art performance in whole-body navigation, manipulation, and non-prehensile pushing for mobile and fixed-base manipulators, outperforming classical MPC and learned-model approaches in accuracy, robustness, and real-world transfer (Pezzato et al., 2023).
Online Adaptive and Data-driven SBPC: Adaptive stochastic sampling with system identification from data provided improved constraint satisfaction, stability, and control performance compared to nonadaptive approaches (Teutsch et al., 2024, Teutsch et al., 2024).

Algorithmic features and sample figures:

Method/Domain	Typical Horizon K	Sample Count/Step	Update Time (ms)	Highlights
Predictive Sampling	23–35	10	4–20	Real-time quad/hand control (Howell et al., 2022)
Dual-spline MPPI	30–40	30–70	20–30	Emergent gaits, CPU only (Schramm et al., 24 Nov 2025)
MPPI (GPU, IsaacGym)	6–15	300–500	50–60	Navigation/manipulation, high DOF (Pezzato et al., 2023)
Trust-region SBPC	40–100	20–300	+15–20% vs MPPI	Smoother, faster convergence (Walker et al., 8 May 2026)
Real hand (Faive)	25	10	<40	0.35 rad/s roll, 67% catch, VLM-tuned (Hess et al., 2024)

6. Extensions, Research Directions, and Theoretical Guarantees

Recent Directions:

Trust-Region and KL-constrained SBPC: KL-bounded updates ensure stability and monotonic cost improvement by avoiding overly aggressive distribution shifts, yielding faster convergence and increased sample efficiency, especially in sample-limited or real-time regimes (Walker et al., 8 May 2026).
Reverse-KL and Accelerated Mirror Descent: Mode-seeking SBPC with negative cost-weighted updates (Reverse-KL), plus Nesterov acceleration and noise-adaptive stepsize heuristics, rapidly finds robust suboptimal policies in under-sampled settings (Kobayashi et al., 2022).
Ising/QUBO SBPC: Binary (QUBO/Ising) representations of the path-integral update, sampling on specialized probabilistic hardware, achieve competitive trajectory optimization with up to 10× fewer samples than Gaussian counterparts (Werthen-Brabants et al., 17 Dec 2025).
Latent-space Learning and Imitation: SBPC controllers trained in learned latent spaces, via end-to-end bi-level optimization, result in higher sample efficiency and improved performance under tight compute constraints (Sacks et al., 2022, Sacks et al., 2022).

Theoretical Guarantees:

Recursive Feasibility and Stability: Suboptimal SBPC updates maintain recursive feasibility and cost improvement, under standard assumptions—warm-starts and terminal constraints/feedback policies serve as practical safeguards (Bobiti et al., 2017, Lorenzen et al., 2016).
Robust Chance Constraint Satisfaction: Sample-based tightening (offline or online) with appropriate sample complexity guarantees that feasible solutions satisfy chance constraints with high probability or confidence (Lorenzen et al., 2016, Teutsch et al., 2024, Teutsch et al., 2024).
Stochastic Lyapunov Analysis: Under mild assumptions, SBPC admits stochastic Lyapunov functions ensuring asymptotic stability in expectation (Lorenzen et al., 2016, Teutsch et al., 2024).

7. Software, Benchmarks, and Deployment Considerations

Research code platforms such as Judo (Li et al., 20 Jun 2025) and MuJoCo-MPC (Howell et al., 2022) provide reference implementations of common SBPC algorithms (Predictive Sampling, CEM, MPPI), flexible controller-optimizer-task abstractions, and GUI/hardware integration interfaces, facilitating rapid prototyping, tuning, benchmarking, and deployment in both simulation and sim-to-real scenarios. These packages support standard tasks (cartpole, manipulation, locomotion), live parameter tuning, and seamless transition to hardware.

Key practical lessons from benchmarking and real deployments:

Exploratory bandwidth and smoothness are traded via noise covariance, spline basis dimension K, and low-pass filter cutoffs.
Sample allocation should match available parallel compute resources; 10–1000 samples per step suffices depending on control horizon and task complexity.
Automated tuning (learning-based weighting, VLM-guided cost adaptation, autotuned temperature) and structure in sampling/parameterization provide consistent empirical gains in performance, efficiency, and robustness.

References: