AMR · IROS 2026

MSC: Multi-Stage Caster-aware Control

Uncertainty-aware NMPC and an RL-based execution layer for caster-equipped industrial mobile robots

Jinho Park^*, Jeongtae Lee^*, Jeonghyeok Lim^*, Keun Ha Choi^†, and Kyung-Soo Kim^†

^* Equal contribution. ^† Corresponding authors.

IROS 2026CU-MPCReinforcement LearningReal-Robot Validation

Overall MSC framework combining CU-MPC, recurrent caster-state estimation, and an RL execution layer — Overall MSC architecture. CU-MPC generates local body-velocity references, while the recurrent execution layer estimates and predicts hidden caster states and maps the references to executable wheel-speed commands.

Why Caster-Aware Control?

Passive caster wheels are mechanically simple and well suited to high-payload industrial AMRs, but their swivel direction is not directly actuated. During low-speed turns, in-place rotations, high-curvature cornering, and forward–reverse transitions, caster misalignment produces lateral scrub forces and resistive reaction torque. The result is a planning–execution gap: a velocity command that is feasible in the planner may be difficult for the drivetrain to realize physically.

Existing solutions typically handle this problem at only one level. Low-level compensation reacts after a command has already been planned, whereas caster-aware planning can become unreliable when caster-state estimates are delayed or uncertain. MSC addresses both issues jointly by coordinating uncertainty-aware planning with contact-consistent low-level execution.

Core Contributions

Integrated Hierarchy

Planning, latent caster-state inference, short-horizon prediction, and wheel-level execution are coupled in one control loop rather than treated as independent modules.

CU-MPC

Predictive caster uncertainty adjusts the caster-alignment cost and input regularization over the NMPC horizon, reducing reliance on uncertain state predictions.

RL Execution Layer

A recurrent actor–critic policy transforms high-level linear and angular velocity references into dynamically executable left/right wheel-speed targets.

Online Control Pipeline

Sensing: acquire robot state and a temporal history of measurable observations.
High-level planning: CU-MPC computes the local reference command (v, ω).
Caster inference: a GRU encoder estimates the current left/right caster states and predicts their short-horizon distributions.
Risk-aware scheduling: predicted variance lowers the caster-aware weight and increases input regularization in uncertain transition regimes.
Execution: the actor maps the high-level command to left/right wheel-speed targets, which are tracked by PI motor controllers.

NMPC horizon63

Control period0.05 s

History length16

Latent dimension48

Prediction horizon8 steps

Project Videos

Short experiment preview. The clip starts automatically in muted mode and repeats continuously.

Full IROS 2026 project explanation video, including the motivation, architecture, and validation results.

Simulation Evaluation

The Isaac Sim model matches the differential-drive and passive-caster configuration of the real robot. Evaluation is performed on three distinct M/S/C trajectories containing low-speed turns, in-place rotation, and high-curvature segments. The comparison isolates both levels of the hierarchy:

B1CA-MPC + kinematics-based PI execution

B2CU-MPC + concurrent-learning execution layer

MSCCU-MPC + proposed recurrent RL execution layer

Trajectory tracking comparison on the M, S, and C paths — M/S/C trajectory tracking comparison. MSC improves the smoothness and consistency of executed trajectories, particularly in turning segments affected by caster contact nonlinearities.

Caster estimation RMSE0.029 / 0.065 / 0.064 radM / S / C paths

Angular-velocity RMSE0.029 / 0.032 / 0.024 rad/sLowest across all three paths

Caster misalignment RMSE0.108 / 0.101 / 0.106 radLowest across M / S / C

Uncertainty-Aware Scheduling

The learned predictor outputs both the expected future caster trajectory and its variance. During rapid caster swing, standstill-to-motion transitions, or other singularity-prone conditions, the predicted risk increases. CU-MPC then reduces the influence of the caster-alignment term and strengthens control regularization, preventing uncertain caster estimates from driving abrupt or torque-sensitive plans.

Caster-state prediction and CU-MPC uncertainty-aware weight scheduling — Learned caster estimation and prediction during a low-speed in-place rotation. Around the rapid caster transition, predictive uncertainty rises; the caster-aware weight decreases while the control regularization weight increases.

Compared with the kinematics-based estimator, the learned estimator reduced caster-angle RMSE from 0.301 to 0.029 rad on the M path, from 0.289 to 0.065 rad on the S path, and from 0.390 to 0.064 rad on the C path.

Real-World Platform

Real-robot validation uses a differential-drive AMR with two freely swiveling caster wheels. Caster encoders provide ground truth only for evaluation. The onboard system includes a LattePanda Sigma computer, motor drivers, a 2D LiDAR, wheel encoders, and an MTi-630R IMU. The navigation software runs on ROS 2 Humble and Nav2.

A 2D map is generated offline using a SLAMTEC S3 LiDAR and Cartographer. During operation, wheel-encoder and IMU measurements are fused by an EKF, and AMCL refines localization against the map. This setup evaluates whether the learned execution layer remains effective with filtering delay, friction variation, micro-slip, and real caster contact effects.

Real AMR hardware platform and ROS 2 navigation stack visualization — Real-world experimental setup: the caster-equipped AMR hardware platform and ROS 2 navigation visualization showing localization, global path, MPC trajectory, and LiDAR measurements.

Real-Robot Results

Real-world C-path tracking, caster estimation, and velocity tracking results — C-path real-robot evaluation. The proposed estimator follows encoder ground truth with less phase lag, while the execution layer improves path and angular-velocity tracking relative to the compared controllers.

Path RMSE0.1163 mBest among compared methods

Angular-velocity RMSE0.0645 rad/sBest among compared methods

Caster estimation RMSE0.1005 radEncoder ground-truth comparison

Misalignment RMSE0.0849 radLowest among compared methods

The improved command realization requires greater actuation effort: mean current increased to 0.9731 A. This tradeoff is consistent with the controller applying the additional effort needed to follow the requested motion under caster-induced resistance.

Key Takeaway

Caster effects are not only a disturbance-rejection problem and not only a planning problem. The experiments indicate that the planning–execution gap is reduced most reliably when uncertainty-aware planning and contact-consistent command realization are designed together. MSC therefore treats caster-state uncertainty as information shared across the hierarchy rather than an isolated estimation output.

Related Publication

MSC: Multi-Stage Caster-aware Control Framework with Uncertainty-Aware NMPC and RL-based Execution Layer →

J. Park, J. Lee, J. Lim, K. H. Choi, and K.-S. Kim, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026.