MuSHR-Type Robots for Autonomous Mobility

• MuSHR-type robots are open-source, low-cost racecars with Ackermann steering designed for multi-agent autonomous research and educational applications.
• They integrate modular hardware, ROS-based software, and advanced sensors to support real-time motion, perception, and coordinated maneuvers.
• The platform offers extensibility in perception, planning, and control, serving as a benchmark for autonomous driving and multi-robot systems.

MuSHR-type robots are open-source, low-cost Ackermann-steered robotic racecars developed primarily by the Personal Robotics Lab at the University of Washington for educational and research applications in multi-agent systems and autonomous mobility. The canonical MuSHR (“Multi-agent System for non-Holonomic Racing”) platform is designed to provide modular hardware, ROS-based software, and detailed open instruction for rapid deployment in both classroom and research settings. While optimized for real-time mobile autonomy and collaborative behaviors at the 1/10-scale, MuSHR-type robots have also been repurposed as robust, contact-rich nonprehensile manipulators in coordinated multi-robot rearrangement systems (e.g., PuSHR). Due to their standardized chassis, sensing, and compute configuration, MuSHR platforms offer extensibility for perception, planning, and control, serving as a widely adopted benchmark for research in mobile robotics, multi-robot coordination, and autonomous vehicle experimentation (Srinivasa et al., 2019, Talia et al., 2023).

1. Hardware Architecture and Platform Characteristics

The standard MuSHR robot adopts a 1/10-scale RC car chassis (e.g., Redcat Racing Blackout SC), characterized by rear-wheel drive and Ackermann steering geometry. This configuration facilitates a non-holonomic, car-like kinematic structure optimized for high-speed maneuvers and precise trajectory following:

• Drive and Steering: A brushed or brushless DC motor actuates the rear axle, while an off-the-shelf hobby servo controls the front steering rack. The mechanical steering range is typically ±30° but can be electronically clamped (e.g., ±10°) for enhanced stability in manipulation tasks (Talia et al., 2023).
• Sensing: Onboard sensors include an Intel RealSense D435i RGB-D camera (depth + IMU), a 2D LIDAR (e.g., YDLIDAR X4), wheel encoders, an IMU, and optional bumper switches for contact detection (Srinivasa et al., 2019).
• Computation: A Jetson-class SBC (e.g., NVIDIA Jetson Nano or TX2) executes the perception, localization, and control stacks. The compute unit is supported by dual NiMH or LiPo batteries, affording run times in the range of ~30–45 minutes depending on payload and operating conditions.
• Networking: Standard Wi-Fi modules enable ROS topic exchange across multiple agents and central workstations.
• Modifications for Manipulation: For nonprehensile pushing tasks (e.g., PuSHR), a 2 cm-thick polycarbonate bumper with a high-friction (μ≈0.6) sandpaper contact surface is mounted, and servo limits are dynamically adjusted during object contact to reinforce quasistatic stability (Talia et al., 2023).

This hardware stack allows the MuSHR platform to support the full pipeline of autonomous perception, planning, and control, with optional expansion for advanced sensors or custom actuation.

2. Kinematics, Dynamics, and Control Framework

MuSHR-type robots are modeled as single-track (bicycle-model) vehicles with the configuration

$p_k = [x_k, y_k, \theta_k]^T \in SE(2),$

and control vector

$$u_k = [v_k, \phi_k]^T,\quad v_k \in [-v_{max}, v_{max}],\quad \phi_k \in [-\phi_{max}, \phi_{max}]$$

at time index $k$. The discrete-time update is

$$\begin{aligned} x_{k+1} &= x_k + v_k \cos \theta_k \Delta t \ y_{k+1} &= y_k + v_k \sin \theta_k \Delta t \ \theta_{k+1} &= \theta_k + \frac{v_k}{L}\tan \phi_k \Delta t \end{aligned}$$

where $L$ is the wheelbase. When integrated with nonprehensile manipulation tasks, $\phi_{max}$ is clamped to $10^\circ$ ($0.17$ rad) to guarantee that the manipulated object remains within the friction-cone constraint during pushing (Talia et al., 2023).

The standard software stack includes several closed-loop control strategies:

• PID Velocity Control
• Pure Pursuit for Path Tracking
• Linear Quadratic Regulation (LQR) for Linearized Models
• Model Predictive Control (MPC/RHC)—the default in advanced deployments—solves a finite-horizon optimal control problem, minimizing deviations from reference trajectories subject to kinematic, dynamic, and collision-avoidance constraints (Srinivasa et al., 2019).

For multi-robot tasks (e.g., PuSHR), each agent locally solves an MPC problem augmented with collision avoidance, consuming local and peer state feedback over Wi-Fi and operating under receding-horizon planning (Talia et al., 2023).

3. Task Assignment, Planning, and Multi-Robot Integration

MuSHR-type robots have been central to compositional approaches to coordinated mobile manipulation and path planning, including the PuSHR multirobot rearrangement paradigm (Talia et al., 2023). System operation typically decomposes as follows:

• Task Assignment (Offline): The rearrangement domain is represented as a grid graph $G=(V,E)$. A mixed-integer program assigns objects to robots so as to minimize total travel and trajectory length, using binary variables $x_{ij}$ to denote robot–object assignments. The Enhanced CBS-TA (Conflict-Based Search–Task Assignment) paradigm is employed to suboptimally solve assignments under user-specified $$u_k = [v_k, \phi_k]^T,\quad v_k \in [-v_{max}, v_{max}],\quad \phi_k \in [-\phi_{max}, \phi_{max}]$$0 scaling.
• Trajectory Generation (Offline): For each assignment, a continuous-space kinodynamic planner (e.g., CL-CBS) computes collision-free, dynamically feasible reference paths $$u_k = [v_k, \phi_k]^T,\quad v_k \in [-v_{max}, v_{max}],\quad \phi_k \in [-\phi_{max}, \phi_{max}]$$1 for every robot-object pair, obeying kinematic constraints ($$u_k = [v_k, \phi_k]^T,\quad v_k \in [-v_{max}, v_{max}],\quad \phi_k \in [-\phi_{max}, \phi_{max}]$$2 rad during manipulation).
• Online Decentralized Control: Each MuSHR runs its own MPC-based tracking controller, using both its own and its peers’ real-time state. Trajectory corrections based on inter-agent distances and modeling uncertainties are performed without global replanning, reinforcing the tractability and robustness of the system.

This architecture avoids the combinatorial intractability of joint assignment and path planning while supporting scalable, real-world multi-robot deployments. Empirically, the PuSHR system achieves near-100% task completion in simulation across diverse scenarios (up to four cars/blocks), with minimum pairwise inter-robot distances exceeding 0.7 m for the majority of the trial duration (Talia et al., 2023).

4. Software, Extensibility, and Open-Source Infrastructure

MuSHR-type platforms run a modular open-source software stack centered on Ubuntu (typically 18.04 LTS) and ROS (Melodic or Noetic). Core functionality includes:

• Perception: Drivers and nodes for RGB-D (RealSense), 2D LIDAR (YDLIDAR), and wheel odometry, producing standardized ROS topics for downstream localization and mapping.
• Localization: Monte Carlo particle filters fusing scan, odometry, and static map data to produce pose estimates, typically on the /amcl_pose topic.
• Planning and Control: Teleoperation modules, autonomous receding-horizon planners (e.g., mushr_rhc), safety controllers, and command multiplexers managing arbitration between control sources.
• Visualization and Experimentation: Real-time visualization via RViz, parametrized launch scripts (e.g., auto.launch), and extensive example Python scripts for direct teleop and Ackermann control.
• Multi-Agent Extension: ROS namespaces and network architectures support the parallel operation of heterogeneous teams, including inter-agent trajectory sharing for decentralized coordination and collision avoidance (Srinivasa et al., 2019, Talia et al., 2023).

Extensibility includes support for advanced perception (e.g., deep learning inference on Jetson TX2/Nano), custom grippers, and adaptation to nonprehensile manipulation by firmware and hardware modification (Talia et al., 2023).

5. Comparison with Related Robotics Platforms

Several low-cost open-source robots have emerged targeting similar educational and experimental domains, notably OpenRoACH (Wang et al., 2019) and HeRoSwarm (Starks et al., 2022). A comparative summary:

Feature	MuSHR	OpenRoACH	HeRoSwarm
Mobility	Ackermann, ~1.5 m/s	Hexapedal, <0.35 m/s	Diff. drive, 0.28 m/s
Compute	Jetson TX2/Nano	BeagleBone Blue / Pi 3	Pi Zero W/Feather BLE
Sensors	RGB-D, LIDAR, IMU, encoders	IMU, camera, line-scan	IMU, mag, prox., mic
ROS version	Melodic/Noetic	Kinetic	Melodic/Noetic
Cost (USD)	$$u_k = [v_k, \phi_k]^T,\quad v_k \in [-v_{max}, v_{max}],\quad \phi_k \in [-\phi_{max}, \phi_{max}]$$3150	~$100
Strength	High-speed autonomy, payload	Durability, terrain	Swarm scalability

MuSHR’s primary distinction lies in payload capacity, modular compute, and high-speed Ackermann steering, positioning it as the testbed of choice for research in autonomous driving and complex multi-agent rearrangement, whereas platforms like HeRoSwarm excel in low-cost, scalable swarm studies but lack computational and sensor breadth (Starks et al., 2022). OpenRoACH offers a legged alternative with superior surface adaptability but lower payload and computational headroom (Wang et al., 2019).

6. Applications and Research Impact

MuSHR and its derivatives have been foundational for:

• Autonomous Driving Research: Curriculum integration at the University of Washington (CSE 490R/571, EE P545) supports hands-on development in localization, perception, mapping, and control for undergraduate and graduate cohorts (Srinivasa et al., 2019).
• Multi-Robot Coordination: Demonstrations of distributed collision avoidance, decentralized RHC, and peer-to-peer trajectory sharing underline the ROS-based modularity of the system.
• Nonprehensile Rearrangement: PuSHR establishes MuSHR-type robots as reliable platforms for contact-rich, coordinated pushing in constrained spaces, leveraging friction-cone analysis and compositional MPC approaches (Talia et al., 2023).
• Benchmarking: MuSHR’s reproducibility and extensibility have led to widespread adoption across over 30 research institutions, establishing it as a de facto standard for benchmarking in mobile robotics and multi-robot systems (Starks et al., 2022).

A plausible implication is that the mature, open-source nature and hardware–software co-design of MuSHR-type robots accelerate reproducible research in robot autonomy, facilitating both classroom deployment and rigorous experimental validation in real-world multi-robot systems.