CLIS-MPPI: Closed-Loop Model Predictive Path Integral Control With Initial-State Sampling Under Execution Delay

arXiv:2610.21239 · cs.RO · Submitted 2026-10-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "CLIS-MPPI: Closed-Loop Model Predictive Path Integral Control With Initial-State Sampling Under Execution Delay".

Dev: The gist The proposed method CLIS-MPPI explicitly addresses execution-time state mismatch by modeling planning delay as a distribution over plausible execution-time initial states,

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: We talked about what CLIS-MPPI does, but let's go over the summary section of "CLIS-MPPI: Closed-Loop Model Predictive Path Integral Control With Initial-State Sampling Under Execution Delay" a bit more. It really lays out the mechanics they are proposing.

Dev: They explicitly address that execution time state mismatch by modeling planning delay as a distribution over plausible execution-time initial states, performing rollout optimization from sampled initial states, and applying a rollout-based local feedback correction during control execution.

Taro: So it’s basically taking the uncertainty of when things actually happen—the delay—and turning it into an input that shapes the entire planning process through sampling and then refining the output with local feedback.

Rosa: Right. They aren't just assuming a fixed state; they are sampling a distribution of states around where you think you started to see what could go wrong during the planning interval itself.

Dev: For every one of those sampled initial states, they generate perturbed control sequences and run parallel forward simulations to get rollout trajectories and their associated costs.

Taro: That’s the heavy lifting, generating all those possible outcomes in advance so you have a distribution of results before you commit to any single path.

Rosa: Then, they use that resulting weighted distribution of rollouts to build a nominal trajectory, and then they extract a local feedback gain from it in a rollout-based manner.

Dev: This extracted gain is then used in the final control input as a closed-loop combination of the MPPI feedforward update and this state-feedback correction term.

Taro: So it’s not just one method; it’s combining the path integral optimization structure with a sampling approach for initialization and then adding a refinement step based on simulated outcomes during execution.

Rosa: It seems like they're really trying to bridge that gap between the abstract optimization of MPPI and the messy reality of real-time execution.

Dev: That's the central theme, connecting the planning phase with what actually happens when you are driving or operating in a dynamic environment.

The paper's summary: Taro: Now let’s look at what they actually suggest as improvements over standard MPPI. They focus on these three main things: delay-aware initial-state sampling for robust planning, online state feedback correction for execution consistency, and enhanced safety in high-frequency, narrow-passage tasks.

Rosa: The first improvement is that they model the delay as an initial-state uncertainty distribution using a covariance defined as Sigma delay, which allows them to sample both initial states and control perturbations.

Dev: That means you're sampling both the execution time initial state and the control sequence, initializing from x(k) t = x ref t + delta(k) t where delta(k) t has a covariance Sigma init.

Taro: So they are explicitly modeling how much uncertainty there is in the starting state and the control itself, which makes the planning much more informed than just assuming everything is perfect at time zero.

Rosa: The second improvement involves extracting a local state-feedback gain from that weighted rollout distribution by fitting a local linear relation to compute it, using weighted least squares.

Dev: That feedback term is then used in the final control input as a closed-loop combination of the MPPI feedforward update and this state-feedback correction term.

Taro: This mechanism ensures that the applied control command better matches the actual robot state at execution time, which directly leads to significantly lower root mean square errors for joint positions and velocities.

Rosa: And finally, they show how this structure provides enhanced safety in tasks like narrow-passage tasks involving obstacle avoidance.

Dev: The results they showed across five real-world runs on a six-DoF manipulator were pretty telling because standard MPPI collided in two of those runs, whereas CLIS-MPPI remained collision-free in all five runs <ref:2610.21239#pg2>.

The paper's improvements: Rosa: So to wrap up what we’ve heard about this paper called "CLIS-MPPI: Closed-Loop Model Predictive Path Integral Control With Initial-State Sampling Under Execution Delay". It seems like the main takeaway is that they successfully transform execution delay into a distribution of initial states and add a rollout-based local feedback correction for closed-loop execution.

Dev: It’s about making sure the applied control command better matches the actual robot state at execution time, which improves consistency between planning and execution, motion smoothness, and safety in delay-sensitive real-world tasks.

Taro: For me, what this means is that we get a system that handles timing uncertainties gracefully by not just crashing when things go wrong but by actively adjusting its control based on the simulation results.

Rosa: The validation on the six-DoF UR5e manipulator showed improved collision avoidance, state consistency, and motion smoothness compared to standard MPPI <ref:2610.21239#pg2>.

Dev: And they also achieved lower joint-position and joint-velocity root mean square errors under CLIS-MPPI, which is a measurable improvement in tracking accuracy.

Taro: Overall, it shows that this approach can provide better safety and stability in environments where timing sensitivity is a real constraint for the hardware.

Rosa: So we have this paper on CLIS-MPPI, which addresses execution delay by modeling it as an initial-state uncertainty distribution and using rollout-based local feedback to correct the control during execution.

Dev: That’s a lot of work to make sure the system is consistent between planning and actual operation, but they showed that for delay-sensitive tasks, it delivers better safety than standard MPPI.

Taro: It’s a solid contribution because it shows how to build robustness into the control loop itself when dealing with real-world timing imperfections.

Conclusion: Rosa: So we’ve been talking about CLIS-MPPI, which is all about handling execution delay by treating it like an uncertainty distribution for the initial state and using rollout feedback to correct the control during execution.

Dev: Exactly, and what I always look at is that this isn't just theoretical; they are modeling that mismatch as a random variable with a sigma delay, which captures all those things like communication jitter or tracking error.

Taro: I mean, the real test for me is how it handles the world when things go sideways. They showed that standard MPPI would collide in two out of five runs in narrow-passage tasks, but CLIS-MPPI stayed collision-free every single time.

Rosa: That’s the key difference there; it’s not just about planning better; it’s about actually being safe when the timing is messy.

Dev: And that safety comes from that rollout-based feedback extraction, where they fit a local linear relation using weighted least squares to compute a state-feedback correction term.

Taro: So when the system runs, it doesn't just rely on the initial plan; it constantly corrects itself based on what the rollout simulations suggest is happening right now.

Rosa: It really means that for someone operating a robot or even something automated in a tricky physical space, this method makes the motion smoother and much more consistent with reality.

Dev: The numbers they put down are pretty telling too; they reduced joint-position and joint-velocity root mean square errors by a big margin under CLIS-MPPI compared to standard MPPI.

Taro: It shows that when you add this layer of uncertainty modeling and feedback correction, you get tangible improvements in precision, not just theoretical gains.

Rosa: So to recap, CLIS-MPPI takes the problem of execution delay and turns it into a distribution over initial states and applies a local feedback policy to ensure the final control matches the actual state at execution time.

Dev: That’s exactly what they did in "CLIS-MPPI: Closed-Loop Model Predictive Path Integral Control With Initial-State Sampling Under Execution Delay."

Taro: It’s a solid way to build robustness into the control loop for tasks where timing is critical, and it definitely gives us a better safety case for real-world deployment.

Rosa: It’s interesting to see how they combine the path integral structure with this sampling approach, but I wonder what happens when the planning horizon gets much longer or when the uncertainty distribution becomes even more complex.

Dev: That’s a good question; one limitation they mention is that they are modeling the delay as Gaussian noise, so if the actual delay behavior is very non-Gaussian, this model might not capture all of it.

Taro: Right, and that means for future work, we might need to explore how to make that initial-state sampling more sophisticated than just a simple normal distribution.

Rosa: Definitely something we should look into next time we review papers on real-time planning; figuring out how to handle those complex failure modes is what keeps me interested.

Zidong Liu, Xu Chen

Department of Mechanical Engineering, University of Washington

cs.RO

Submitted: 2026-10-15

Updated: 2026-10-15

Comments: 65th IEEE Conference on Decision and Control (CDC 2026)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: The gist The proposed method CLIS-MPPI explicitly addresses execution-time state mismatch by modeling planning delay as a distribution over plausible execution-time initial states, performing rollout

Key concepts

Execution-Time State Mismatch: This occurs when the robot's actual state at the time of control application differs from the state used during planning simulations. CLIS-MPPI models this mismatch as a distribution of plausible initial states around the measured state, accounting for factors like communication jitter and tracking errors.
Initial-State Sampling: Instead of starting optimization from one fixed point, this technique samples a distribution of initial states around the current measured state. This allows the planner to consider various possible system evolutions during the planning interval, leading to more robust control sequences.
Rollout-Based Feedback Correction: After generating multiple rollout trajectories, CLIS-MPPI extracts a local state-feedback gain by fitting a linear relation to the weighted distribution of these rollouts. This gain is then used as a correction term during execution, ensuring the actual control input aligns with the optimized plan.

Summary

The gist The proposed method CLIS-MPPI explicitly addresses execution-time state mismatch by modeling planning delay as a distribution over plausible execution-time initial states, performing rollout optimization from sampled initial states, and applying a rollout-based local feedback correction.

How it works

CLIS-MPPI is designed to compensate for the mismatch between the state used in forward simulation and the actual state at control execution by modeling planning delay as a distribution over plausible execution-time initial states. Instead of assuming that the control optimization starts from a single measured state, CLIS-MPPI samples a distribution of perturbed initial states around the currently measured state to represent the possible system evolution during the planning interval. For each sampled initial state, perturbed control sequences are generated and parallel forward simulations are performed to obtain rollout trajectories and their associated costs.

Key Components of CLIS-MPPI

The proposed method augments standard MPPI through three specific components:

Delay-Induced Initial-State Uncertainty: This component models the execution delay as a random variable, specifically defining the mismatch as a random variable where the mismatch is defined as x apply t - x ref t. The mismatch is modeled as ∆xt ∼ N (0,Σdelay), where Σdelay captures the state uncertainty induced by execution delay, communication jitter, zero-order hold, and lowlevel tracking error.
Initial-State-Sampled MPPI: This involves sampling both the execution-time initial state and the control sequence. For rollout k, it samples δ(k) t ∼ N (0,Σinit) and initializes the rollout from x(k) t = x ref t + δ(k) t.
Rollout-Based Feedback Extraction and Closed-Loop Execution: This involves extracting a local state-feedback gain from the weighted rollout distribution. It fits a local linear relation u˜(k) t+j ≈ Kt+jx˜(k) t+j via weighted least squares to compute the feedback gain. The final control input is then computed as a closed-loop combination of the MPPI feedforward update and a state-feedback correction.

Validation and Results

The effectiveness of CLIS-MPPI was validated through real-world experiments on a 6-DoF manipulator in a narrow-passage task involving both obstacle avoidance and end-effector target reaching. Across 5 real-world runs of the same task, standard MPPI collided in 2 runs, whereas CLIS-MPPI remained collision-free in all 5 runs. Averaged over all collision-free runs for each method (3 for MPPI and 5 for CLIS-MPPI), CLIS-MPPI reduced the average task completion time from 25.4 s to 24.4 s and shortened the average endeffector path length from 0.456 m to 0.425 m. Furthermore, the overall joint-position and joint-velocity root mean square errors (RMSEs) were reduced from 0.100 rad and 0.108 rad/s under MPPI to 0.017 rad and 0.011 rad/s under CLIS-MPPI.

Conclusion

CLIS-MPPI transforms execution delay into an initial-state uncertainty distribution and augments MPPI with a rollout-based local feedback policy for closed-loop execution. This approach effectively improves the safety of MPPI in delaysensitive tasks by preventing collisions that can arise from delay-induced state mismatch, while also producing smoother and more stable robot motion during execution. The effectiveness of the proposed method was validated through real-world experiments on a 6-DoF UR5e manipulator, where CLIS-MPPI demonstrated improved collision avoidance, state consistency, and motion smoothness compared to standard MPPI.

ACKNOWLEDGMENT Support for this research was provided in part by an Amazon Science-Hub gift award.

Improvements for AI systems

Bold header: Delay-Aware Initial-State Sampling for Robust Planning

This improvement directly addresses execution-time state mismatch by modeling it as a distribution over plausible initial states using the covariance defined as Σinit = Σdelay. This allows CLIS-MPPI to sample both initial-state and control perturbations, and computes the applied control input using a closed-loop combination of the MPPI feedforward update and a rollout-based local feedback correction, leading to improved safety.

Bold header: Online State Feedback Correction for Execution Consistency

The system can now apply a state-feedback correction term to the feedforward update, which is computed via weighted least squares fitting a local linear relation derived from rollout deviations. This mechanism ensures that the applied control command better match[es] the actual robot state at execution time, resulting in significantly lower root mean square errors (RMSEs) for joint positions and velocities.

Bold header: Enhanced Safety in High-Frequency, Narrow-Passage Tasks

The CLIS-MPPI system can reliably navigate environments sensitive to timing and state uncertainty, such as narrow-passage task involving obstacle avoidance. The validation showed that standard MPPI collided in 2 runs, whereas CLIS-MPPI remained collision-free in all 5 runs, demonstrating superior safety under execution delays.

Sources

Model Predictive Path Integral Methods with Reach-Avoid Tasks and Control Barrier Functions