Autonomous Driving Small-Scale Cars: A Survey of Recent Development

In plain words

Self-driving research usually needs full-size cars, which are expensive and tightly regulated on public roads. This survey looks at the small robot cars people use instead, from toy-like kits meant for classrooms to research cars that carry real sensors and computers. It sorts them into groups: educational kits, commercial research platforms, open-source builds, outdoor racing cars, and multi-vehicle test cities.

The authors also list what researchers have done with these cars: lane keeping, racing, obstacle avoidance, car following, overtaking, drifting, parking and cooperative driving. Many tasks are split into classic methods and machine-learning methods. The paper ends with suggestions: make platforms easier to start with, give them better computers and sensors, build smart-city test areas, and study car-to-car communication more. These suggestions are the authors' opinion; nothing was tested.

TL;DR

• A survey of small-scale autonomous-car platforms, covering hardware, software, simulators, sensors and driving tasks §1.1.
• The hardware and software tables are compiled from the literature by the authors; the paper runs no experiments §1.1 tab. II tab. III.
• Driving tasks are split into a modular pipeline (perception, planning, control) and an end-to-end pipeline §4.
• The main value for an implementer is the platform list (scale, steering, compute, price) and the per-task method tables tab. II tab. IV–tab. VII.
• Use it to choose a platform or find the methods used on it. Verify every price and specification against the vendor, because the paper gives no dates for them.

Problem

Full-size autonomous-driving research is costly and heavily regulated, which limits small research groups and individuals from testing on public roads §1. The authors say no earlier work gives a full overview of these small platforms, including hardware, software and benchmarked driving tasks. They name two earlier surveys: one that lacks a comparative analysis and one that covers platforms without the techniques used on them §1.1.

Method

This is a literature survey. The authors state that they analysed "over 250 research papers" and placed the driving tasks into two pipelines §1.1 §4. The paper does not describe how the papers were selected.

The structure of the survey, with the section each part is in:

1. Platforms §2. Categories are educational, commercial research, open-source, outdoor and multi-vehicle §2.1.1–§2.1.5. Hardware is compared in tab. II and software, simulators and focus tasks in tab. III.
2. Simulators and sim-to-real §2.2 §2.3. The paper lists general and platform-specific simulators, the two ROS versions, and four transfer methods: zero-shot transfer, transfer learning, domain adaptation and domain randomization.
3. Sensors and compute §3.1–§3.8. Camera, LiDAR, IMU and encoders, GPS, central localization (motion capture or markers), other sensors, compute units and vehicle-to-everything communication.
4. Benchmarking of driving tasks §4. Two pipelines are described fig. 2:
   • Modular: perception (localization and mapping §4.1.1, lane recognition §4.1.2, obstacle recognition §4.1.3, traffic signs §4.1.4), planning §4.2 and control §4.3. Each subsection gives classic and machine-learning methods, with comparison tables tab. IV–tab. VII.
   • End-to-end §4.4: imitation learning (IL) and deep reinforcement learning (DRL) on raw sensor input, mainly for lane keeping and racing.
   • Proposed framework §4.5: an untested baseline pipeline that combines the methods above. See the pseudocode below.
5. Future trends §5 and Conclusion §6.

The paper's own baseline (§4.5), as the authors describe it. The paper gives no parameters, so none are filled in here:

# Modular baseline pipeline, one control step  [§4.5]
# Input:  sensors (camera, LiDAR, IMU, encoders, GPS)
# Output: steering and throttle command

def drive_step(sensors, goal):
    # 1. Fuse the sensors into one state estimate with a DNN or a
    #    Kalman filter, and update the map with SLAM  [§4.5] [§4.1.1]
    state = fuse_sensors(sensors)
    world_map = slam_update(sensors, state)

    # 2. Detect and classify nearby objects with a shallow model,
    #    for example an SVM or k-nearest neighbours  [§4.5]
    objects = shallow_detector(sensors)

    # 3. Plan a global path with A*, Dijkstra or RRT  [§4.5] [§4.2]
    path = global_planner(world_map, goal)

    # 4. Choose the behaviour (lane keeping, lane change, overtaking,
    #    obstacle avoidance) with a state machine, DRL, IL, MPC or a
    #    decision tree  [§4.5]
    behaviour = behaviour_planner(state, objects, path)

    # 5. Turn the behaviour into a command with a tracking controller;
    #    the options are listed in the control section  [§4.3]
    command = track(behaviour, state)

    # 6. Safety layer: fall back to a PID or rule-based stop when a
    #    failure is detected  [§4.5]
    return safety_filter(command, state)

For end-to-end systems the paper says raw sensor data goes "directly" to IL, RL or a hybrid, with the safety module still in place §4.5. That is the same shape as the pseudocode with steps 1–5 replaced by one learned policy.

Hyperparameters

Not applicable: the paper is a literature survey §1.1 and trains and evaluates nothing, so there are no hyperparameters. Platform specifications are reported under Results.

Results

No experiments. The survey's data are the compiled tables below, taken from the paper's own hardware table. The paper gives no date for the prices, and it prints them in USD only. "--" means the paper does not state a value.

Platform overview, condensed from tab. II (the full table also lists sensors and footnote details):

Source for every row: tab. II. "Commercial" is the paper's "commercially available" column. Footnote 7 of tab. II says a build guide with off-the-shelf parts exists for several "no" rows.

Software and focus tasks, from tab. III. Examples: AutoRally, F1TENTH and DeepRacer are listed for racing; Duckietown and Donkeycar for lane keeping; e-puck, Kilobot and Pheeno for swarm work; UDSSC and Chronos for cooperative driving. Simulators in the same table include Gazebo (most ROS platforms), Chrono (ART/ATK), SUMO (UDSSC), LGSVL (CoRoLa) and Webots/V-REP/ARGoS (swarm platforms).

Method comparisons (qualitative, with the platforms each reference used), from tab. IV (localization and mapping), tab. V (obstacle detection), tab. VI (planning) and tab. VII (control). The tables use check and cross marks for pros and cons, not measured values.

Sensor and compute figures stated in the text §3:

• Raspberry Pi: about 32 GFLOPS, 4-core 1.5 GHz CPU §3.7.
• Nvidia Jetson: about 472 GFLOPS; the highest model is "up to 300 TOPS" §3.7.
• Microcontrollers: 100 KFLOPS to 25 MFLOPS §3.7. PCs: 50 GFLOPS (CPU) to several TFLOPS (GPU) §3.7.
• Single-plane RPLiDAR: 12 m range, 0.225° resolution, 10 Hz scan, about 400 USD §3.2. The same section also gives a 40 m range, about 1 % error and 32,000 samples per second for "the RPLiDAR" §3.2 (see Pitfalls).
• Hokuyo UST-10LX: 270° field of view, 10 m range §3.2.
• Hobby LiDAR from about 150 USD; IMUs from about 15 USD; encoders often under 10 USD; GPS about 30 USD; RTK GPS several hundred USD; motion-capture systems above 10,000 USD §3.2 §3.3 §3.4 §3.5.
• Conventional GPS accuracy: about 5 m §3.4.
• Camera: Raspberry Pi camera at 1080p with 44 dB dynamic range §3.1.

Paper vs code

No official code. This is a survey and has no software release. The simulator repositories footnoted in tab. III (gym-donkeycar, gym-duckietown and f1tenth_gym) belong to the platform projects, not to the authors, so they are not linked here. The paper's proposed framework §4.5 has no implementation.

Pitfalls

• No validation. Every comparison in tab. IV–tab. VII is a qualitative summary by the authors. The baseline pipeline in §4.5 is a proposal and is not tested.
• Dated and unstated prices. The paper gives no date for the USD prices and no currency conversion. "--" means not stated, not free.
• Table II Donkeycar runtime. The "Runtime" cell for Donkeycar holds a check mark, which looks like a misplaced entry from the "Commercial" column. The paper's own footnote calls the platform a build guide with off-the-shelf parts, so the "no" in the commercial column is the one to trust.
• Conflicting RPLiDAR figures in one section: range 12 m and 10 Hz in one paragraph, 40 m and about 32,000 samples per second in another §3.2. The paper does not say which model each figure belongs to.
• Mixed Ackermann and differential platforms. The text says that newer platforms moved to Ackermann steering §2; the tables still list many differential-drive platforms, especially educational and swarm ones tab. I tab. II.
• Educational kits are weak for research. The paper says they have closed hardware and software and seldom have networking §2.1.1. This rules them out for multi-vehicle or research work.
• Sim-to-real evidence. The paper names domain randomization as an option in Duckietown and DeepRacer §2.3 and cites sim-to-real studies, but the survey text reports no transfer numbers for a specific platform.
• Localization dependence. Some tasks use indoor motion-capture or marker-based localization §3.5 §4.3.5. The paper says such central systems are expensive and cannot easily be moved, so plan for that if the platform must run outdoors.
• Source contains commented-out material. The LaTeX has commented-out tables (an overview table in §2 and a task overview in §4.4) and a red-highlighted paragraph. These are not part of the compiled paper and are not cited here. The table numbers used in this card follow the order of the active tables.
• Figures are not reproduced. fig. 1 shows platform images taken from their own papers or websites. This card describes it in words and does not copy it.
• Selection of papers. The 250-plus papers are not listed, and the body does not state the selection rule. Treat the coverage as the authors' reading.

Lineage

• Builds on: none stated. The authors contrast it with two prior surveys: Caleffi et al. (2023), which lacks comparative analysis, and Mokhtarian et al. (2024), which covers platforms only §1.1. No arXiv ids are given in the paper, and neither survey has a card here.
• Compared against: the two surveys above, plus the platform papers cited in tab. II and tab. III.
• Superseded by: not known from content in this wiki.