Wheeled Lab: Open Sim2Real Robotics Platform

Updated 14 January 2026

Wheeled Lab is an open-source robotics platform integrating modular hardware, high-fidelity simulation (Isaac Lab), and ROS for reproducible research.
It employs advanced domain randomization and realistic sensor modeling to bridge the simulation-to-reality gap in wheeled robotic systems.
The platform enables diverse RL tasks—controlled drifting, elevation traversal, and visual navigation—with robust experimental performance metrics.

Wheeled Lab is a modern, low-cost, open-source research and education platform that integrates advanced simulation-to-reality (“Sim2Real”) robotics methodologies, control architectures, and datasets for wheeled robotic systems. It provides a comprehensive physical and virtual testbed bridging high-fidelity physics simulation (Isaac Lab), robust perception (LiDAR, cameras, odometry), and learned policies—enabling reproducible, scalable research and democratizing access to cutting-edge robotics techniques through openly accessible hardware and code bases (Han et al., 11 Feb 2025).

1. System Architecture: Hardware, Simulation, and Interfaces

Wheeled Lab comprises modular, extensible hardware (standard 1/10-scale RC platforms and MuSHR-type robots), integrated with NVIDIA Isaac Lab for simulation and ROS for real-world deployment. Key hardware features include brushed DC motors for drive, a front-wheel steering servo (in drift mode), 3D LiDAR, RGB cameras, visual-inertial odometry (VIO) modules, and compute via NVIDIA Jetson Orin NX or Odroid-class systems. The HOUND platform is representative: wheelbase 0.28 m, track width 0.23 m, 3.5 kg mass, sub-0.05 m center of mass, and full autonomy package cost ≈ $3000. The stripped-down MuSHR variant costs ≈$930 (Han et al., 11 Feb 2025).

The software stack centers on Isaac Lab, providing high-fidelity agent physics (multi-DOF vehicles with collision contacts and damped suspension), physically realistic sensor models (LiDAR, cameras, IMUs, VIO drift), and massive parallel training of agents (64–1024 concurrent simulations per GPU). Domain randomization modules perturb physical properties (friction, mass, actuator gains, lighting), while a highly modular RL pipeline interfaces with both simulation and ROS, enabling direct policy transfer (Han et al., 11 Feb 2025).

2. Simulation-to-Reality Transfer: Dynamics, Sensing, and Domain Randomization

Wheeled Lab prioritizes Sim2Real generality by combining detailed modeling and domain perturbation. Vehicle dynamics are modeled using both the standard bicycle model and per-wheel contact physics; tire friction is drawn from μ ∼ Uniform(0.2, 0.8), vehicle mass and actuator delays are randomized per roll-out, and sensor observations are augmented with Gaussian noise and vision distortions. For perception, IMU signals (σ_acc ≈ 0.02 m/s², σ_gyro ≈ 0.001 rad/s), camera frames (resolution 40×60, ±30% jitter/blurring), and 16-beam LiDAR scans are corrupted to bridge the terrain and sensor distribution gap between sim and real (Han et al., 11 Feb 2025).

Open-loop calibrations (e.g. spring-scale drag for friction, actuator gain tuning) and suspension parameter identification further close the simulation gap. Reward function shaping, observation dropout, and visual/physical texture randomization address the principal failure modes typical of unstabilized transfer (Han et al., 11 Feb 2025).

3. Policy Learning: Reinforcement Learning, Architectures, and Tasks

The framework incorporates state-of-the-art RL methods (proximal policy optimization; PPO) and policy architectures (MLP and CNN). Training occurs over thousands of epochs with aggressive environment parallelization, e.g., 64 envs × 5000 epochs for drifting, 1024 envs × 5000 epochs for elevation traversal (Han et al., 11 Feb 2025). Core observation spaces include vehicle kinematics, onboard state, and environment representations (elevation mapping, image features).

Task policies demonstrated in Wheeled Lab include:

Controlled drifting (π_drift): minimizes cross-track error and maintains forward velocity on tight ovals, incentivizing slip-angle and turn-energy. Policies generalize to maintain control up to 58° slip angle with robust recovery from spinout.
Elevation traversal (π_elev): navigates ramped and obstacle-rich terrains using elevation maps, attaining 90% success rate in Sim2Real transfer, maintaining ascent speed ≈ 0.5 m/s and max roll <15°.
Visual navigation (π_vis): CNN and MLP-based policies for line/path following under vision augmentations. Sim2Real transfer successful primarily with strong augmentation; MLP with color jitter and blur achieves up to 60% real-world success (Han et al., 11 Feb 2025).

4. Quantitative Performance and Experimental Results

Performance is systematically reported via task-completion rates, trajectory tracking metrics, inference latency, and safety margins. Drift policies achieve 100% real-world lap completeness at 1.6 m/s average, with policy inference ≈5 ms on embedded compute. Elevation policies traverse physical ramps/obstacles with 9/10 success, tightly bounding dynamic state. Vision policies' Sim2Real efficacy is contingent on domain adaptation techniques; without augmentation, transfer fails on all trials (Han et al., 11 Feb 2025).

Relative to existing open wheeled frameworks, Wheeled Lab expands beyond standard MuSHR/F1Tenth systems by integrating Sim2Real policy transfer, full physical and sensor domain randomization, and generalized task-learning pipelines. It complements and is interoperable with recent research in advanced wheeled mobility (e.g., Verti-Wheelers for vertical terrain (Datar et al., 2023), data-driven dynamics D4W (Lin et al., 2024), and omnidirectional or reconfigurable mobile bases (Zhao et al., 2024)). The system also supports integration with hybrid wheeled-legged robots and hierarchical RL navigation policies as in (Lee et al., 2024), but at a fraction of the hardware cost and with a lower technical barrier to entry (Han et al., 11 Feb 2025).

6. Open Source, Educational Impact, and Future Extensions

Wheeled Lab is fully open source, with hardware and software resources hosted at https://github.com/UWRobotLearning/WheeledLab. Ready-to-use Isaac Lab environments, ROS wrappers, CAD/URDF files, and comprehensive guides support rapid onboarding for research and teaching. The platform facilitates undergraduate/graduate lab exercises in control, RL, and Sim2Real, and supports competitive events (e.g., drift racing, maze navigation). Planned extensions include open curricula, community task modules, online adaptation tools, and systematized curriculum learning for friction estimation and auto-tuning (Han et al., 11 Feb 2025).

A key implication is that Wheeled Lab serves not only as a low-cost research testbed, but also as a pedagogical bridge, democratizing access to advanced wheeled robotics and enabling rapid, robust evaluation of autonomous control and perception algorithms under diverse, realistic conditions.