⚡ electrical engineering

Smooth Sampling-Based Model Predictive Control Using Deterministic Samples

This paper proposes deterministic sampling MPPI (dsMPPI), a novel control framework that combines the exponential weighting of MPPI with deterministic sampling and cross-entropy method optimizations to generate smoother trajectories for nonlinear systems compared to existing state-of-the-art methods.

Original authors: Markus Walker, Marcel Reith-Braun, Tai Hoang, Gerhard Neumann, Uwe D. Hanebeck

Published 2026-06-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Markus Walker, Marcel Reith-Braun, Tai Hoang, Gerhard Neumann, Uwe D. Hanebeck

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to drive a truck into a tight parking spot or balance a broom on its handle. To do this, the robot uses a smart planning tool called Model Predictive Control (MPC). Think of this tool as a "what-if" simulator. Before the robot moves, it runs thousands of mental simulations, asking: "If I turn the wheel this way, then that way, will I succeed?" It picks the best plan and executes the first step, then repeats the process.

The problem with many current simulators is that they use random guessing to create these "what-if" scenarios. It's like throwing darts blindfolded at a board to find the best path. While this works, the resulting instructions can be jittery and chaotic. The robot might jerk the steering wheel left, then right, then left again in rapid succession. In the real world, this "chattering" is bad news; it wears out motors and makes for a bumpy ride.

The New Solution: "dsMPPI"

The authors of this paper propose a new method called dsMPPI (Deterministic Sampling Model Predictive Path Integral Control). Here is how it works, using simple analogies:

1. From Random Darts to a Perfect Grid
Instead of throwing darts randomly (which leaves big empty gaps and clumps of darts in one spot), the new method uses deterministic sampling. Imagine instead of throwing darts, you lay out a perfect, evenly spaced grid of points to cover the entire board.

The Benefit: This ensures the robot explores every possible option evenly without missing spots or wasting time checking the same spot twice. It's like sweeping a floor with a perfectly organized broom rather than flailing it randomly.

2. The "Soft" vs. "Hard" Selection
The paper combines two existing ideas:

The Old Way (CEM): This method picks the "top 10" best plans from its simulations and ignores everything else. It's like a teacher who only looks at the top 10 test scores and throws away the rest. This can be too harsh and lead to jerky decisions.
The New Way (MPPI style): This method looks at all the plans but gives more weight to the good ones and a little bit of credit to the okay ones. It's a "soft" selection.
The Mix: The new dsMPPI uses the perfect grid (deterministic sampling) combined with this "soft" weighting. It gets the best of both worlds: a thorough search that doesn't ignore good ideas, leading to much smoother instructions.

3. The "Permutation" Trick
To make sure the robot doesn't get stuck in a rut (since the grid is always the same), the authors added a clever trick called permutation.

The Analogy: Imagine you have a deck of cards representing different parts of the robot's movement. In one round, you look at the cards in order. In the next round, you shuffle the deck and look at them in a different order.
The Benefit: This keeps the search fresh and diverse without needing to store massive amounts of extra data. It's like rearranging the furniture in a room to see if a new layout works better, without buying new furniture.

What Did They Find?

The researchers tested this new method on two classic challenges:

Swinging a pole up: Getting a pendulum to stand straight up on a moving cart.
Backing a truck: Reversing a long trailer into a parking spot.

The Results:

Smoother Rides: The new method produced control inputs (steering, acceleration) that were significantly smoother than the old random methods. The robot didn't jerk around; it moved gracefully.
No Extra Cost: Even though the math is more complex, the computer didn't take any longer to solve the problem. It was just as fast as the random methods.
Better Performance: In many cases, it found better solutions (lower costs) than the random methods, especially when the robot had to make many quick decisions.

The Bottom Line

The paper claims that by swapping random guessing for a smart, organized grid of possibilities, and by mixing in a "soft" way of choosing the best plans, robots can move much more smoothly. This means less wear and tear on the robot's parts and no need for extra "filtering" software to smooth out the jerky movements after the fact. It's a more efficient, gentler way for robots to learn how to move.

Technical Summary: Smooth Sampling-Based Model Predictive Control Using Deterministic Samples

Problem Statement

Sampling-based Model Predictive Control (MPC) has gained prominence for its ability to handle complex nonlinear systems and nonconvex cost functions without relying on gradient-based optimization. However, a critical limitation of existing methods, such as the Cross-Entropy Method (CEM) and Model Predictive Path Integral (MPPI) control, is the generation of non-smooth control inputs. This issue arises because the update of proposal distribution parameters is approximated using a finite number of random samples, leading to noisy estimates. In practical applications, such noisy inputs can cause excessive actuator wear, rapid velocity reversals, and torque switching. While post-processing techniques like low-pass filtering are commonly used to mitigate this, they result in suboptimal performance because the smoothing is not integrated into the optimization process.

Methodology

The paper proposes deterministic sampling MPPI (dsMPPI), a novel algorithm that integrates the strengths of CEM and MPPI with deterministic sampling to produce smoother control trajectories. The core components of dsMPPI include:

Deterministic Sampling: Instead of random sampling, the method utilizes deterministic samples generated offline by minimizing the modified Cramér–von Mises (CvM) distance. These samples are transformed online to match the current proposal distribution parameters (mean and covariance). This ensures low discrepancy, covering the search space without clustering or gaps.
Exponential Weighting (MPPI-style): Unlike standard CEM, which relies on a "hard" elite set selection (equally weighting the top $N$ samples), dsMPPI employs a soft, exponential weighting scheme based on sample costs: $w^{(i)} \propto \exp(-J(\bar{\xi}^{(i)})/\lambda)$ . This allows for smoother updates to the proposal distribution.
Iterative Optimization: The algorithm performs multiple iterations ( $j_{max}$ ) per time step to refine the proposal distribution, similar to improved CEM (iCEM), rather than the single-step update typical of standard MPPI.
Momentum Smoothing: To stabilize optimization and prevent premature convergence, momentum smoothing is applied to both the mean and covariance updates.
Adaptive Temperature: An inverse temperature parameter ( $\lambda$ ) is adapted heuristically at each iteration based on the normalization constant of the weights, ensuring robust performance without manual tuning.
Variation Schemes for Exploration: To prevent poor exploration when reusing deterministic samples, two variation schemes are introduced:
- Multi-iteration method: Selecting different subsets from a larger pre-calculated pool of samples for each iteration.
- Permutation method: Randomly permuting the dimensions of the deterministic samples in every iteration. This approach requires storing only a single set of samples while promoting diversity through dimension shuffling.
Temporal Correlations: To further enhance smoothness, the covariance matrix is constructed using a fixed time-correlation matrix derived from colored noise (power spectral density law), updating only marginal variances to maintain computational efficiency.

Key Contributions

dsMPPI Algorithm: The integration of deterministic sampling with MPPI's exponential weighting and CEM's iterative framework.
Permutation Variation Scheme: A memory-efficient method to enhance exploration with deterministic samples by shuffling dimensions, offering a trade-off between cost performance and smoothness compared to the multi-iteration subset selection.
Comprehensive Evaluation: Extensive simulations on two nonlinear benchmark tasks: the cart-pole swing-up and the truck backer-upper.

Experimental Results

The proposed dsMPPI was evaluated against standard MPPI, dsCEM (deterministic sampling CEM), and an iterative random-sampling variant (MPPI Iterative).

Smoothness: dsMPPI consistently achieved the smoothest control inputs across both tasks. In the cart-pole swing-up, it outperformed both dsCEM and MPPI Iterative, particularly in "settled smoothness" (the second half of the trajectory), which is critical for avoiding oscillations near the set point. Visualizations confirmed significantly reduced chattering compared to standard MPPI.
Cost Performance: While dsCEM generally achieved slightly lower cumulative costs than dsMPPI, dsMPPI consistently outperformed its random-sampling counterpart (MPPI Iterative), demonstrating the benefits of deterministic sampling. The permutation scheme showed a slight cost advantage over the multi-iteration scheme, particularly at lower sample sizes.
Computational Efficiency: Standard MPPI was the fastest due to its non-iterative nature. However, all iterative methods (dsMPPI, dsCEM, MPPI Iterative) exhibited similar computation times. Crucially, the use of deterministic sampling incurred no additional online computational overhead compared to random sampling.

Significance and Claims

The paper claims that dsMPPI effectively leverages the exponential weighting of MPPI alongside the efficiency and low-discrepancy properties of deterministic samples. The primary significance lies in the ability to generate significantly smoother control trajectories without the need for suboptimal post-processing filtering. This reduction in mechanical stress and actuator wear makes the method particularly suitable for real-world applications. Furthermore, the authors emphasize that these improvements are achieved without increasing the online computational burden, making dsMPPI a compelling upgrade for existing sampling-based MPC implementations. The permutation variation scheme is highlighted as a novel contribution that balances exploration and memory efficiency.