TEACar: An Open-Source Autonomous Driving Platform

1. Problem Statement

Intelligent Transportation Systems (ITS) research increasingly relies on vision-based perception and learning-based control, necessitating experimental platforms that support realistic hardware-in-the-loop validation. While simulation environments (e.g., Gazebo) are useful for early development, they fail to capture the complexity and uncertainty of physical systems.
Existing small-scale autonomous platforms face significant limitations:

• RoboRacer (F1/10): While robust, its 1/10 scale is large, expensive (~$3,000), and tightly coupled to specific chassis, limiting flexibility for vision-based tasks and reconfiguration.
• DonkeyCar: While accessible and vision-focused, it suffers from non-modular software design, reliance on separate power sources for computation and actuation (increasing integration complexity), and limited extensibility due to its custom Python stack.

There is a need for a compact, modular, and cost-effective platform that decouples mechanical, hardware, and software subsystems to facilitate rapid prototyping and rigorous validation of learning-based control algorithms.

2. Methodology and System Design

The authors propose TEACAR, a 1/14- to 1/16-scale autonomous racing platform designed around three core pillars: Compactness, Modularity, and Reliability.

A. Mechanical Design

The platform utilizes a four-layer deck architecture to physically decouple subsystems, improving rigidity and simplifying reconfiguration:

1. Chassis Adapter Interface Layer: Bridges the commercial RC chassis and the deck. It is interchangeable to accommodate different mounting geometries (hole spacing, fasteners) without altering the rest of the deck.
2. Lower Structural Base: Houses the battery, maintains a low center of gravity, and manages reaction forces from the suspension.
3. Middle Electronics Layer: Provides a rigid mounting surface for computing units (e.g., NVIDIA Jetson), power distribution boards, motor drivers, and IMUs, ensuring thermal clearance and consistent alignment.
4. Upper Structural Frame: Supports forward-facing sensors (cameras) and antennas, elevating them to improve the field of view and providing mechanical protection during rollovers.

Key Advantage: When switching chassis, only the adapter layer requires redesign; the remaining deck layers remain invariant.

B. Hardware Design

• Computing Unit: Supports various NVIDIA Jetson modules (Xavier NX, Orin NX, Orin Nano), offering 21–100 TOPS of AI performance.
• Power System: Features a custom Power Distribution Board (PDB) that manages varying voltage requirements (9–15V for compute/motor, regulated 5V/7V for sensors/servos) and includes protection circuits against over-voltage/under-voltage.
• Sensors & Actuators: Supports CSI/USB cameras, IMUs (MPU6050/9250), and joysticks. It uses a PCA9685-based PWM controller to interface with the computing unit via I2C, providing up to 16 independent channels for precise motor and servo control.

C. Software Design

The software stack is built on ROS 2 (Humble) to ensure modularity and real-time performance. It is organized into three subsystems:

1. Controller: Manages inputs from joysticks or trained Neural Networks (CNNs). It publishes a standardized motion_cmd message containing steering and throttle targets.
2. Actuator Driver: A two-layer architecture. The first layer aggregates commands from controllers and converts them to PWM pulse widths. The second layer interfaces with the PCA9685 hardware via I2C to generate physical signals.
3. Perception: Handles sensor data acquisition (e.g., gscam for images, joy for controllers) and publishes data in standardized ROS 2 message types.

3. Key Contributions

• Modular Architecture: A novel four-layer deck design that physically decouples sensing, computation, actuation, and power, allowing for easy reconfiguration and chassis swapping.
• Hardware Abstraction: A standardized power distribution and actuation interface that supports diverse RC chassis and computing units without requiring custom wiring for every new setup.
• Open-Source Ecosystem: The complete design files, software stack, and documentation are available on GitHub, lowering the barrier to entry for researchers and educators.
• Cost-Effectiveness: The platform costs approximately $1,200, significantly less than competitors like RoboRacer (~$3,000), while maintaining high computational capabilities.

4. Experimental Results

The authors constructed a prototype and evaluated it across mechanical, hardware, and software dimensions.

• Mechanical Evaluation:

  • Dimensions: Total mass of 1.25 kg (vs. 3.13 kg for RoboRacer) and a minimum turning radius of 550 mm (vs. 1200 mm), making it highly suitable for constrained lab environments.
  • Rigidity: Survived three 1-meter drop tests with no structural damage or functional degradation.

• Hardware & Software Performance:

  • Inference Latency: Three CNN-based steering controllers ("Small," "Medium," "Large") were tested. The average inference latency ranged from 1.8 ms to 2.8 ms, well within real-time requirements.
  • Power Consumption: The computing unit consumed an average of 7.5–7.6 W, well below its 25W rated limit, allowing for sustained operation.
  • Comparison: When compared to a DonkeyCar platform (using a Jetson Nano), TEACAR's inference latency was roughly 6 times faster (e.g., 1.8 ms vs. 12.4 ms for the "Small" model), attributed to the superior compute power of the Jetson Orin NX.

5. Significance

TEACAR addresses the critical gap between simulation and full-scale vehicle testing by providing a scalable, modular, and affordable testbed.

• For Research: It enables rapid iteration of vision-based perception and learning-based control algorithms in a realistic physical environment without the prohibitive costs of full-scale vehicles.
• For Education: Its modular design allows students to easily understand and modify individual subsystems (mechanical, electrical, software), fostering deeper learning in autonomous systems.
• Future Impact: The platform is designed to be extensible, with future work planned to integrate LiDAR and depth cameras for advanced sensor fusion and mapping tasks.

In conclusion, TEACAR represents a significant step forward in democratizing autonomous driving research, offering a robust alternative to existing platforms by prioritizing modularity and hardware abstraction.