IEEE Transactions on Robotics Vol. 41, 2025 ICRA 2026 Poster Presentation

Single-Instance Sampling for Computationally Efficient and Accurate Real-Time Task Space MPPI Control

Dongwhan Kim^*1, Euncheol Im^*2,3, Yujin Kim⁴, Myotaeg Lim³, Yisoo Lee^2,†

* Co-first authors † Corresponding author

¹LG CTO, ²Korea Institute of Science and Technology (KIST), ³Korea University, ⁴Cornell University

D. Kim and Y. Kim were formerly with KIST and Korea University

Paper Video

The proposed method enables 1 kHz real-time task space control of a 7-DoF robot manipulator using single-instance sampling MPPI with an average computation time of only 0.298 ms.

0.298 ms

Average computation time
(fastest MPC-based task space controller)

18×

Faster than STORM
(4.62 ms → 0.298 ms)

17×

Faster than FDDP
(17.14 ms → 0.984 ms)

50×

Fewer samples than
conventional MPPI

Abstract

This study presents a model predictive path integral (MPPI) method capable of conducting high-frequency real-time model predictive control (MPC) for robot manipulators. Real-time MPC-based manipulation holds significant potential for controlling an end-effector precisely and reactively while satisfying various constraints in dynamic environments.

However, the optimization under a complex robot model and various constraints imposes a heavy computational burden, hindering the realization of high-frequency updates. To address this challenge, we propose a single-instance sampling-based MPPI algorithm and dynamic time horizon to significantly reduce the computational burden while enhancing control performance.

The performance and efficacy of the proposed method are verified through experiments conducted on a 7-degree-of-freedom robotic arm (Franka FR3), along with comparative simulations and analysis against STORM, cuRobo, and FDDP.

Proposed Method

Overview of the proposed MPPI-based task-space control framework. Yellow blocks execute in parallel on GPU.

Single-Instance Sampling

We propose a novel sampling strategy where a constant change in control input (δu) is sampled once and applied throughout the entire predictive horizon for each trajectory.

This approach differs fundamentally from conventional MPPI, which samples different control perturbations at each time step, requiring nested for-loop computations.

Key advantages:

Eliminates nested for-loops → replaces with single parallel GPU execution
Broader trajectory exploration without increasing sample count
Achieves comparable/better performance with 50× fewer samples
Computation time for sample generation: only 4.09 × 10⁻³ ms

Dynamic Time Horizon

We propose an adaptive time horizon that automatically adjusts its length based on the current distance to the goal.

The prediction horizon is divided into two segments using binary segmentation: a short fixed-step segment (dt₁ = 0.001 s) for near-future precision, and a variable-step segment (dt₂) for long-range exploration.

Key advantages:

Predicts up to 2.55 s into the future with only 64 time steps
Short horizon when near goal → fast convergence
Long horizon when far from goal → avoids local minima
Automatically resolves joint limit saturation and local minima scenarios

Experiments

Validated on a 7-DoF Franka FR3 robotic arm with real-world experiments and MuJoCo simulations.

Control Performance

In point-to-point real-world experiments, the proposed method achieved:

Average position error: ~0.001 m
Average orientation error: ~0.018 rad
Average computation time: 0.298 ms (enabling >1 kHz control)

Sample generation and state prediction together take only 3% of total computation time.

Point-to-point control: position & orientation converging to 6-DoF target.

Constraint Handling

Joint Limit Avoidance

Under a 5π/3 rad yaw rotation command, joint 7 reaches its position limit and maintains that position. Joints 1 and 2 then generate larger movements to continue tracking the target orientation without violating joint constraints.

Singularity & Workspace Boundary Avoidance

When the target is set outside the reachable workspace, the robot moves to the closest feasible position while the manipulability-based cost keeps it above 0.02 — ensuring no loss of controllability.

Self-Collision Avoidance

Using a neural network-based collision cost, the robot detects and avoids self-collision in real time, achieving the closest feasible posture to the target without explicit trajectory planning.

Comparison

vs. STORM

18× faster computation (0.255 ms vs. 4.62 ms)

12× more accurate pose tracking (L2-norm: 3.09×10⁻² vs. 3.80×10⁻¹)

STORM fails to converge orientation error to zero; proposed method fully converges.

vs. FDDP

17× faster computation (0.984 ms vs. 17.14 ms)

Comparable tracking accuracy with longer prediction horizon (up to 2.55 s vs. 0.9 s for FDDP).

Position error: 0.005 m (proposed) vs. 0.035 m (FDDP).

vs. Spline Sampling

1.46× faster computation (0.255 ms vs. 0.372 ms)

Single-instance sampling requires only one sample point per rollout, vs. at least two nodes for spline interpolation.

Comparison of position and orientation errors: proposed method vs. STORM (default, N=500 K=30) vs. STORM (same parameters, N=128 K=64).

Comparison among STORM and cuRobo

Sequential target tracking across 9 poses including out-of-reach and self-collision targets (simulation).

The proposed method and STORM handle out-of-reach / self-collision targets gracefully. cuRobo fails with "IK solve fail" and cannot respond to dynamically updated targets.

Comparison with FDDP

Real-world experiment on 7-DoF Franka FR3.

17× faster computation

0.984 ms (proposed) vs. 17.14 ms (FDDP)

Better position accuracy

0.005 m (proposed) vs. 0.035 m (FDDP)

Longer prediction horizon

up to 2.55 s (proposed) vs. 0.9 s (FDDP)

BibTeX

@article{kim2025single,
  author    = {Kim, Dongwhan and Im, Euncheol and Kim, Yujin and Lim, Myotaeg and Lee, Yisoo},
  title     = {Single-Instance Sampling for Computationally Efficient and Accurate Real-Time Task Space {MPPI} Control},
  journal   = {IEEE Transactions on Robotics},
  volume    = {41},
  pages     = {6327--6344},
  year      = {2025},
  doi       = {10.1109/TRO.2025.3626660},
}