💻 computer science

STR Robot: Design of an Autonomous Mobile Robot from Simulation to Reality

This paper presents a simulation-to-real implementation of an autonomous mobile robot that focuses on developing onboard control, self-localization, and navigation systems, demonstrating the effectiveness of simulation as a foundation for reliable real-world deployment.

Original authors: Vinh Nguyen, Gia-Uy Le, Tien-Dat Nguyen, Tri-Tin Nguyen, Vinh-Hao Nguyen

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vinh Nguyen, Gia-Uy Le, Tien-Dat Nguyen, Tri-Tin Nguyen, Vinh-Hao Nguyen

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 teaching a new driver how to navigate a busy city. You wouldn't just throw them into rush hour traffic immediately; that would be dangerous and expensive if they crashed. Instead, you'd start them in a video game.

This paper is about building a self-driving robot car (specifically one that steers like a real car, with a front wheel that turns, rather than a tank that spins in place) and using that exact same "video game" approach to get it ready for the real world.

Here is the story of how they did it, broken down into simple parts:

1. The "Video Game" Training Ground (Simulation)

The team started by building their robot inside a computer simulation (using a tool called Gazebo). Think of this as a flight simulator for pilots.

Why? It's safe, cheap, and you can crash a thousand times without breaking anything.
The Goal: They wanted to see if the robot's "brain" could learn to find its way around obstacles and drive smoothly before ever touching a real machine.

2. The Robot's Three Brains

To drive itself, the robot needs three main skills, which the authors built into a single system:

The GPS (Localization & Mapping): The robot needs to know where it is. It uses a spinning laser scanner (LiDAR), a camera, and an internal gyroscope (IMU) to build a 3D map of the world around it. It's like the robot is constantly taking photos and measuring distances to figure out, "I am here, and that wall is over there."
The Navigator (Path Planning): Once it knows where it is, it needs a route. They used a smart version of the "A*" algorithm (a classic math trick for finding the shortest path).
- The Upgrade: Standard planners often make paths that are too sharp or hug obstacles too closely. The team improved this so the robot plans smoother, safer routes that give it plenty of room to turn, like a careful driver who doesn't cut corners.
The Driver (Path Tracking): This is the part that actually turns the steering wheel. They compared two "drivers":
1. The Standard Driver: A common method that tries to follow the line but sometimes gets jittery or cuts corners too aggressively.
2. The New Driver (A-GMPC): This is the team's special invention. It uses advanced math to understand the robot's physical shape (Ackermann steering). It's like a driver who not only looks at the road but also feels the steering wheel, ensuring the turns are smooth and the car doesn't jerk around.

3. From the Game to the Real World (Sim-to-Real)

The big challenge in robotics is that what works in a video game often fails in real life because real sensors are noisy and the ground isn't perfectly flat.

The Test: They took the exact same code they wrote in the simulation and put it on a physical robot with real lasers and cameras.
The Result: It worked! The robot successfully drove itself outdoors.
- In the "video game," the new driver (A-GMPC) was much more accurate than the standard one, especially on tricky shapes like figure-eights.
- In the real world, the new driver was also superior. It stayed closer to the intended path, especially when making sharp turns, whereas the standard driver tended to drift or overshoot.

4. The "Magic" of the New Driver

Why was the new driver (A-GMPC) better?
Imagine you are driving a car. If you jerk the steering wheel left and right rapidly, the car feels unstable. The standard computer driver sometimes did this mathematically. The new driver was programmed to be "smooth." It penalized itself for making jerky steering moves, forcing it to find a way to turn that felt natural and gentle, just like a human driver would.

Summary

The paper claims that they successfully built a self-driving robot system that:

Learns in a simulation first to save time and money.
Transfers perfectly to reality, meaning the code didn't need to be rewritten for the real robot.
Drives smoother and safer than older methods, thanks to their new "geometric" steering controller.

They even made all their code and robot designs public so other people can use them to build their own robots. The paper concludes that this "simulation first" approach is a reliable way to build robots that can navigate the real world safely.

Technical Summary: STR Robot: Design of an Autonomous Mobile Robot from Simulation to Reality

Problem Statement
The development of autonomous mobile robots for outdoor environments faces significant challenges, including sensing uncertainty, model mismatches, and limited computational resources. While simulation tools offer a safe and efficient environment for prototyping, transferring navigation frameworks from simulation to real-world deployment remains difficult due to discrepancies in sensing and environmental complexity. Furthermore, existing navigation systems often treat localization, planning, and control as separate modules rather than a unified onboard system. There is a specific need for a practical, reproducible "simulation-to-real" pipeline that addresses these integration challenges, particularly for non-holonomic platforms like Ackermann-steered vehicles, which require smooth trajectory tracking to avoid oscillatory behavior or corner-cutting.

Methodology
The paper proposes a complete simulation-to-real autonomous navigation framework built upon an existing mechanical platform. The system is structured into three primary modules: onboard localization and mapping, global path planning, and path tracking.

Onboard Localization and Mapping:
The system employs a dual-layer estimation architecture. A local Extended Kalman Filter (EKF) fuses high-rate wheel encoder and IMU data with lower-rate scan-based odometry (LiDAR-Inertial or LiDAR-Inertial-Visual) to provide smooth, high-frequency state feedback for the controller. At the global level, a second estimator corrects the local odom frame against a 2D occupancy map using ICP-based scan-to-map matching. For mapping, the system utilizes FAST-LIVO2 for normal illumination and FAST-LIO2 as a fallback for low-light conditions, generating a registered 3D point cloud that is projected into a 2D occupancy grid for ground navigation.
Path Planning:
The global planner utilizes an improved A* algorithm on a 2D grid map. To ensure safety, obstacles are expanded based on the vehicle's dimensions. The cost function is modified to penalize direction changes, encouraging smoother paths. Post-processing removes redundant inflection points to further smooth the trajectory. During execution, a local obstacle avoidance module dynamically updates the reference path using C2-smooth curves (Bezier curves) to navigate around moving obstacles while preserving the global route.
Path Tracking (A-GMPC):
The core innovation lies in the path tracking controller. The authors compare a standard Model Predictive Control (MPC) baseline (using linearized Euclidean error states) with a proposed Ackermann Geometric MPC (A-GMPC).
- Geometric Formulation: A-GMPC defines tracking errors directly on the $SE(2)$ manifold rather than in Euclidean space, improving spatial consistency.
- Input-Increment Optimization: To address the physical constraints of Ackermann steering and prevent aggressive control actions, the controller optimizes the input increment ( $\Delta u_k$ ) rather than the absolute input. This penalizes rapid changes in steering and velocity, promoting smoother actuation.
- Implementation: The controller solves a finite-horizon quadratic optimization problem subject to actuator bounds, converting the optimized linear/angular velocity commands into Ackermann steering angles ( $\delta$ ) via the kinematic bicycle model.

Key Contributions
The paper outlines three primary contributions:

A Complete Simulation-to-Real Pipeline: The authors present a unified framework integrating onboard localization, path planning, and path tracking, validated first in Gazebo simulation and subsequently on a physical robot.
A-GMPC Integration: The work introduces an Ackermann-oriented Geometric MPC controller into a unified autonomy stack. This controller leverages $SE(2)$ geometric error formulations and input-increment optimization to enhance tracking smoothness and accuracy for non-holonomic platforms.
Open-Source Release: The experimental setup, source code, and results are made available to the community to facilitate reproducibility and further research.

Results
The framework was evaluated through both simulation and real-world experiments involving lemniscate and square trajectories at various speeds (0.6 m/s to 1.5 m/s).

Path Planning: The improved A* algorithm generated paths that were shorter (22.35m vs. 25.52m) and significantly smoother (170.55° total turning angle vs. 3712.50°) compared to the baseline, albeit with a slight increase in planning time.
Path Tracking (Simulation): A-GMPC demonstrated lower Root Mean Square Error (RMSE) than the standard MPC on both trajectories, with a notable improvement on the lemniscate path (0.0764m vs. 0.1215m).
Path Tracking (Real-World): In physical experiments, A-GMPC consistently outperformed the standard MPC across all tested speeds and trajectories. For instance, at 1.5 m/s on a square trajectory, A-GMPC achieved an RMSE of 0.6182m compared to 1.1475m for the standard MPC. The results indicate that the geometric formulation and input-increment optimization effectively mitigate tracking errors during sharp maneuvers.
Runtime: Both controllers operated within the required 30 Hz control loop (33.3 ms), confirming real-time feasibility.

Significance and Claims
The paper claims that the proposed approach demonstrates the feasibility of a robust simulation-to-real deployment for autonomous mobile robots. By integrating modern planning and tracking modules into a unified onboard system, the work establishes a practical pipeline that bridges the gap between simulation and reality. The authors emphasize that simulation provides an effective foundation for developing reliable systems, provided that the control strategies (like A-GMPC) are designed to handle the specific kinematic constraints and actuator limitations of physical hardware. The work does not claim to solve all challenges in dynamic outdoor navigation but provides a validated framework for smooth, safe, and autonomous navigation on Ackermann-steered platforms. Future work is identified as focusing on robust tracking of humans and moving objects in more complex environments.