CPM Lab: Cyber-Physical Mobility Research

• Cyber-Physical Mobility Lab is an open-source platform that integrates physical testbeds, simulators, and digital twins for rigorous cyber-physical mobility research.
• It features a four-layer architecture from high-level controllers to low-level microcontrollers, ensuring deterministic timing and reproducible sim-to-real experiments.
• The lab advances learning-based control, multi-agent coordination, and formal safety verification to enhance autonomous vehicle and smart city transport research.

The Cyber-Physical Mobility Lab (CPM Lab) is an open-source research infrastructure for the rigorous development, validation, and benchmarking of autonomous and networked mobility systems. It integrates microrobotic vehicles, simulators, middleware, and hardware-in-the-loop architectures for joint algorithmic and experimental exploration of cyber-physical mobility problems, with a particular emphasis on the intersection of learning-based control, formal safety, and multi-agent coordination. Foundational application domains include connected and automated vehicles (CAVs), cooperative driving, smart city transport systems, and cyber-physical drone operations (Kloock et al., 2020, Beerwerth et al., 23 Jan 2026, Malikopoulos, 2021).

1. Core Hardware and Software Architectures

CPM Lab is organized as a four-layer architecture, each with dedicated roles:

1. High-Level Controllers (HLC): Scenario coordination, routing, behavior/trajectory planning, collision avoidance, and formal safety verification. HLCs run on external Intel NUCs, supporting both centralized and distributed operation modes. Trajectory plans are communicated via a pub/sub middleware (Kloock et al., 2020).
2. Middleware (DDS + Logical Execution Time): Implements OMG-DDS publish/subscribe mechanisms (RTI Connext DDS), and overlays logical execution time (LET) scheduling guaranteeing deterministic timing. Inputs, computations, and outputs are aligned to fixed logical cycles, eliminating race conditions and enabling reproducibility (Kloock et al., 2020):

$$\text{Sense: at } kT_\text{cycle} \to \text{Compute: } kT_\text{cycle}+\Delta_\text{sense} \to \text{Actuate: } kT_\text{cycle}+\Delta_\text{sense}+\Delta_\text{comp}$$

1. Mid-Level Controllers (MLC): Onboard Raspberry Pi Zero W units for each μCar, handling trajectory tracking, model predictive control (using Cubic Hermite interpolation for reference reconstruction), sensor fusion from hall-effect odometers, IMUs, and external position feeds.
2. Low-Level Controllers (LLC): On ATmega2560 microcontrollers. Sample sensors at 1 kHz, actuate motor/steer commands via PWM, and enforce low-level safety (Kloock et al., 2020).

All layers are replicated in simulation, permitting seamless code reuse between simulated and hardware-in-the-loop experiments.

Physical agents are 1:18-scale μCars (220 × 107 × 70 mm, ≤3.7 m/s), equipped with a full sensor suite and external indoor positioning by overhead cameras. Full construction plans, source code, and build instructions are open-source (Kloock et al., 2020, Beerwerth et al., 23 Jan 2026).

2. Integrated Simulation and Digital Twin Pipeline

CPM Lab supports a structured, reproducible sim-to-real pipeline:

Domain	Vehicle Model	Control Stack	State Acquisition	Timing/Sync
Simulation	Kinematic bicycle model	Direct policy input	Gazebo/world	ROS2 + LET
Digital Twin	Grey-box vehicle log fitting	Mid-level MPC	Sensor noise, delays	ROS2 + LET
Physical Lab	1:18 μCars, onboard PID/MPC	Mid-level MPC	Overhead camera, Pi	ROS2 + LET

The integration ensures policies (e.g., SigmaRL MAPPO-based multi-agent RL) are deployed bit-identically across the pipeline, only diverging in terms of control interface and physical realism (Beerwerth et al., 23 Jan 2026). This system supports rapid prototyping, transfer learning studies, and evaluation of architectural and environmental mismatches, essential for sim-to-real research.

3. Algorithmic Frameworks: Learning, Control, and Decentralized Coordination

CPM Lab research advances mathematical and algorithmic methods for cyber-physical mobility:

• Controlled Markov Frameworks: System models employ controlled Markov chains ($x_{t+1} \sim P(\cdot|x_t, u_t)$), Bellman optimality equations, decentralized information structures, and constraint sets for physical safety (distance buffers, control limits) (Malikopoulos, 2021).
• Model Predictive and Learning-Based Control:
  • Powertrain Learning (POD): Online update of transition statistics and cost estimates, with state-action Q-value minimization and stochastic approximation guarantees.
  • CAV Intersection Control: Each vehicle solves a local optimal control problem (OCP), minimizing $\int a_i(\tau)^2 d\tau$ subject to linear dynamics, boundary conditions, safety constraints, yielding cubic-polynomial trajectories tracked with MPC.
  • Safety-Critical RL/CBF-QP: Integration of RL policy outputs with control barrier function optimization ($u^* = \arg\min_u\|u-u_{nom}\|^2$, s.t. $L_f h(x) + L_g h(x) u + \alpha h(x) \geq 0$), guaranteeing collision-free operation while preserving learning performance (Malikopoulos, 2021).
  • Situation-Aware DRL for Drones: Uses DDPG in Unity-based cyber-physical simulation, with state augmented by K-raycast obstacle measures for nonlinear trajectory adaptation. In dense scenarios, DDPG + Raycast yielded up to +356% performance boost versus linear baseline (Lee et al., 2022).
• Multi-Agent RL and Zero-Shot Transfer: MARL approaches (MAPPO, SigmaRL) operate with decentralized actors and centralized critics, trained in simulation and deployed zero-shot in digital twin and physical testbed, enabling structured evaluation of sim-to-real gaps (Beerwerth et al., 23 Jan 2026).

4. Experimental Platforms and Benchmark Scenarios

CPM Lab has demonstrated a range of application-driven experimental studies:

• Platooning and Multi-Vehicle Coordination: Automated leader-follower control, distributed MPC, and stress testing with up to 20 μCars. Lateral tracking errors <50 mm, zero collisions at cycle rates of 100 ms (Kloock et al., 2020).
• Intersection Management and Conflict-Free Traversal: Priority-based inter-vehicle arbitration (Alrifaee), distributed scenario coordination, measured throughput ~3 vehicles/min, average conflict wait times ~1.2 s (Kloock et al., 2020).
• Sim-to-Real MARL Benchmarking: Zero-shot deployment of MARL policies, collision rate rises linearly from simulation → digital twin → physical lab (CRA-A: 0.37 → 2.10 → 4.49 events/100 m). Structured metrics include cumulative reward, RMSE of trajectories, and success rate $\rho$ (Beerwerth et al., 23 Jan 2026).
• Self-Learning Powertrain and PHEV Sizing: Experimental powertrain demos recorded 5–10% fuel savings and 8% NOₓ reduction; PHEV sizing with post-transmission motor placement yielded 12% economy gain (Malikopoulos, 2021).
• Autonomous Drone Mobility: Situation-aware DDPG achieves successful mission rates of 11.9/20 episodes in high-obstacle environments, versus 6.1/20 for linear control (Lee et al., 2022).
• Scaled Smart City Testbeds: 1:25 scale urban networks (UDSSC), decentralized optimal control eliminates stop-and-go, achieving ≈25% average travel time reduction and observed energy implications (Beaver et al., 2019, Malikopoulos, 2021).
• Indoor Multi-Vehicle Wireless Testbed: Four agent vehicles with Raspberry Pi neural compute, multi-modal sensors, and particle filter-based IPS; closed-loop spacing regulation robust to up to 30% packet loss (Bemani et al., 2020).

5. Formal Methods, Monitoring, and Safety Assurance

CPM Lab contributes formalism for spatio-temporal verification of distributed mobility systems:

• Spatio-Temporal Logic (STREL): Extends Signal Temporal Logic with spatial modalities (reach, escape). Allows system monitoring by defining logical properties over space and time (e.g., connectivity, battery reliability), supports constraint semiring semantics and invariance under Euclidean isometries (Bartocci et al., 2019).
• Cognitive-Inspired Distributed Club Detection: Mobile agents use local state vectors, Markov-normalized adjacency, and inflation phases for detecting dynamic, overlapping social communities via physical encounter logs (HCMM), with accuracy >95% under low mixing (Massaro et al., 2013).
• Safety via Control Barrier Functions: Forward-invariant safe sets maintained via quadratic program constraints impose strict separation ($\|p_j-p_i\| \geq \Delta_{safe}$), enabling provable collision avoidance in CAV and mixed autonomy traffic (Malikopoulos, 2021).

6. Scalability, Reproducibility, and Open Science

CPM Lab is entirely open-source, offering public access to code, hardware blueprints, reproducibility guides, and scalable architecture:

• Code/Build Resources: All assets available at https://cpm.embedded.rwth-aachen.de, with step-by-step deployment instructions for HLC/MLC middleware, microcontroller firmware, and track assembly (Kloock et al., 2020).
• Remote Access: Planned web-based interface allows authenticated scenario upload, live monitoring, and per-user sandboxed experiments with formal security controls.
• Co-Simulation and Mixed-Agent Integration: Physical vehicles and unlimited simulated agents are integrated seamlessly via middleware, maintaining logical synchrony and determinism across experimental modes.
• Extension and Adaptability: CPM Lab supports rapid extension to new vehicles by launching additional HLC+MLC pairs, integration of alternate sensors (IPS, BLE), digital twin feedback, and incorporation of novel communication technologies (5G, DSRC) (Bemani et al., 2020).

7. Open Problems and Future Research Directions

Priority avenues for future exploration include:

• Partial Penetration and Mixed Traffic: Developing adaptive methods for aggregation and modeling of human-driven versus autonomous vehicle interactions.
• Scalability: Extending corridor and intersection control frameworks to city-scale, multi-intersection networks, with coupled dynamics.
• Incorporating Stochastic Human Behavior: Modeling uncertainty via POMDPs, trust learning, and partial observability for robust policy design.
• Social Metrics and Equity: Integrating mechanism design and socially-aware control objectives to account for acceptance, fairness, and policy impact.
• Resilience: Engineering robust operation under communication failures, delays, cyberattacks, or agent faults; buffer-based and predictive control provide avenues for graceful degradation.
• Unified Learning and Control: Advancing approximate information-state methods to unify computational efficiency, safety, and learning for real-time, distributed mobility systems (Malikopoulos, 2021, Beerwerth et al., 23 Jan 2026).

The CPM Lab thereby establishes a reproducible, rigorously instrumented foundation for the advancement and empirical assessment of cyber-physical mobility systems, encompassing all stages from theory and simulation to deployment and formal assurance.