Autonomous Vehicles on the Edge: A Survey on Autonomous Vehicle Racing

In plain words

Racing a car with no driver is a harder problem than ordinary self-driving. The car has to go as fast as it can, close to the limit where the tyres stop gripping, and it has to find other cars, pass them, and avoid crashing into them. This paper is a map of the research done on that problem up to the end of 2021.

It sorts the work the same way a self-driving stack is usually built. First the car has to know where it is and what is around it (perception). Then it has to choose a fast line and a safe path around other cars (planning). Then it has to steer and brake to follow that path (control). A last group of papers tries to learn the whole chain with one neural network (end-to-end). The authors also describe the racing cars that researchers can actually use, from small remote-controlled cars to full-size race cars, and the competitions they race in.

The paper ends with a list of open problems: seeing objects reliably at high speed, planning with several cars at once, modelling tyres well enough to use in real time, and balancing safety against speed. It contains no new method and no experiments of its own, so its main use is as a reading list and a vocabulary for the field.

TL;DR

A survey, not a method paper. It covers autonomous racing cars and reviews the surveyed work in four software areas (perception, planning, control, end-to-end) plus applied studies, and the hardware platforms and competitions §1 §2 §3. It covers 233 papers published up to the end of 2021 §5. Planning is split into global (minimum lap time or minimum curvature lines), local (obstacle-avoiding trajectories) and behavioural (overtaking decisions) §2.2; model predictive control is the most common control approach §2.3. Nothing is re-implemented or benchmarked here. Numbers in the Results section are those the survey reports for other papers.

Problem

Autonomous racing is the setting where a car must drive at the limits of handling, at high speed, and against adversaries, with the human driver replaced by software §1. The authors argue that road-driving stacks cannot be transplanted directly, so racing acts as a testbed whose methods are later moved to passenger cars §1.

The survey defines its scope narrowly: four-wheeled cars, combustion or electric, real or small-scale, that either run on a racetrack, race against others, or solve a racing problem. Drone racing and high-speed freeway driving are excluded because neither the limit-handling nor the adversarial aspect is present §1.2.1. The survey claims to be the first holistic review of the field covering both software and hardware §1.1.

Each software area comes with its own list of fundamental problems: high-speed object detection and localisation without HD maps §2.1; minimum-time global lines, long local planning horizons, real-time replanning and interaction with other cars §2.2; accurate path and speed tracking, and exact modelling of nonlinear vehicle behaviour §2.3; and the data and generalisation demands of end-to-end systems §2.4.

Method

The survey's "method" is its taxonomy. The categorisation follows a perception, planning, control pipeline §1.2.2 fig. 1. The compiled paper is structured as follows: §1 introduction, §2 autonomous racing software (§2.1 perception, §2.2 planning, §2.3 control, §2.4 end-to-end, §2.5 applied studies), §3 hardware and competitions (§3.1 small-scale vehicles, §3.2 real race vehicles), §4 open research questions, §5 summary.

Receding-horizon MPC loop §2.3 fig. 7

The survey's general description of MPC, which most controllers in §2.3 follow. The loop is written in our own words.

# Receding-horizon control loop  [§2.3]
# Input:  reference trajectory with a speed profile, horizon N
# Output: steering and acceleration commands, one per step

def control_loop(reference, horizon):
    # Re-plan at every step from the newly measured state  [§2.3]
    while driving:
        state = measure_vehicle_state()
        # Solve the finite-time optimal control problem over N steps  [§2.3]
        states, inputs = solve_ftocp(state, reference, horizon)
        # Apply only the first input, then repeat at the next time step  [§2.3]
        apply(inputs[0])

Perception §2.1 tab. 1 fig. 2 fig. 3

• Localisation dominates. Most perception work is localisation: SLAM or Monte Carlo (AMCL) localisation on LiDAR, extended Kalman or H-infinity filters that fuse odometry, IMU and GPS, and graph-SLAM with cone detection on Formula Student cars §2.1 tab. 1.
• Mapping and free space. Google Cartographer and GMapping are adapted to large outdoor racetracks, and a camera-and-depth DNN detects free space where no lane lines exist §2.1.
• Object detection on Formula Student cones is based on YOLO variants, SSD MobileNet and a shared dataset. The survey notes that object detection at racing speeds has received little attention §2.1.
• No HD maps. None of the listed perception papers uses pre-built HD maps. An open-source set of about 30 racetracks gives only inner and outer bounds, so it can support planning but not localisation §2.1.

Planning §2.2 tab. 2 fig. 4 fig. 5 fig. 6

Planning is split into three layers. Each one is summarised with its method families as the survey groups them.

• Global planning computes a raceline, the fastest path around the whole track, usually by minimising lap time §2.2. Four families are named:
  • Evolutionary algorithms (e.g. CMA-ES, particle swarm), where each individual encodes a racetrack and vehicle configuration and is scored by its simulated lap time §2.2.
  • Optimal control problems with lap time as the objective, solved with SQP, Newton's method or gradient-based nonlinear programming, subject to track geometry and friction limits §2.2.
  • Minimum-curvature lines, where the racecar's dynamic asymmetry is ignored, which the survey calls not ideal for a car that is asymmetric in braking and acceleration §2.2.
  • Geometric lines that imitate the decision points of a human driver (braking, turn-in, apex, turn-out, acceleration points) and interpolate them with splines or clothoids §2.2.
• Local planning plans a short, fixed horizon that avoids other cars while staying near the global line §2.2. Three strategies are named: modify the global plan by optimisation (usually MPC); sample several dynamically feasible motion primitives and select the best by cost; or sample free space for goal states (RRT and variants, which give no optimality guarantee) §2.2.
• Behavioural planning makes high-level choices, such as overtaking left or right, staying behind, energy management, and reacting to race control flags. Two strategies are named: weighted multi-cost selection, and game-theoretic methods that treat the race as a step-by-step game and use regret or best-response ideas §2.2.

Control §2.3 tab. 3 fig. 7 fig. 8 fig. 9

Control turns a reference trajectory (position and speed profile) into steering and acceleration commands. The aim is low lateral and heading error, low speed error, and stable behaviour at high acceleration §2.3. The control actions are steering angle and an acceleration request sent to low-level actuator controllers §2.3.

• Classic control. Feedback and feedforward steering, pure pursuit and lookahead controllers, controllers that operate on a predefined G-G diagram, centre-of-percussion steering, and slip-angle control of the front tyres §2.3.
• Model predictive control (MPC), described as the most popular strategy §2.3. The variants named are: nonlinear MPC (solved with FORCES Pro on the vehicle); hierarchical NMPC with a point-mass high-level and a high-fidelity low-level model; LPV-MPC; tube MPC, which bounds the effect of disturbances with a tube of reachable sets; and sampling-based MPC (MPPI, and robust MPPI). Sparse randomised and stochastic MPC use data collected along the track to tighten constraints §2.3 fig. 8.
• Optimisation-based tuning. TUNERCAR tunes both control gains and vehicle hardware parameters (centre of gravity, mass) with CMA-ES, using lap time as the score §2.3.
• Learning-based control. Iterative learning control (ILC) adds a feedforward correction from earlier laps §2.3 fig. 9. Learning MPC (LMPC) uses the trajectories of earlier laps to make each lap faster while keeping constraints §2.3. Gaussian-process models learn the gap between the planned and the driven path, online or in an NMPC §2.3.
• Drifting control. Controllers that hold a steady high-slip-angle drift while following a reference path, such as a circle or figure of eight §2.3.

End-to-end §2.4 tab. 4 fig. 10

End-to-end systems replace some or all modules with a DNN. Partial end-to-end keeps a classic controller on top of a learned intermediate representation (e.g. a trajectory); full end-to-end predicts the actuator commands directly §2.4. The survey calls racing a good proving ground because the driveable area is clear and the objective, lap time, is clear. It names the main issues as system architecture, learning nonlinear tyre dynamics, the simulation-to-reality gap, data volume, and rare high-speed events §2.4.

Applied studies §2.5 tab. 5 fig. 11 fig. 12

Simulation environments (TORCS, the Roborace Learn-to-Race Gym interface, the SVL simulator, F1TENTH Gym, Formula Student and the Roborace Metaverse), vehicle dynamics modelling, complete software stacks, and evaluation studies §2.5.

Hardware and competitions §3 tab. 6

• Small-scale 1:10 cars (F1TENTH, the Berkeley car, MIT Racecar, MuSHR, RoSCAR) with Jetson-class GPUs and interchangeable cameras and LiDARs. F1TENTH races a time trial and a head-to-head knockout race every year §3.1.
• 1:43 cars in the ETH ORCA testbed: a vision system tracks the cars, a separate control platform computes MPC inputs, and they are sent to the cars over Bluetooth §3.1.
• AutoRally, a 1:5 scale RC truck with two cameras and RTK GPS, used for planning and control on low-friction surfaces §3.1.
• Real race vehicles: Roborace (Devbot 2.0, a Le Mans prototype chassis with a fully electric drivetrain, with the Roborace Season Beta series), the Indy Autonomous Challenge (Indy Lights chassis, combustion engine, target speed 290 km/h), eV Grand Prix Autonomous (electric go-karts, 1:3 scale) and Formula Student Driverless (teams build their own driverless car) §3.1 §3.2 tab. 6.

Open research questions §4

Nine challenges are listed §4:

1. High-speed perception: no work covers high-speed detection, a detection range above 100 m is needed, and there is no public high-speed dataset §4.
2. Multi-vehicle trajectory planning with more than three cars, non-convex and recursively feasible, executable in real time §4.
3. Multi-vehicle interaction: prediction with multimodal, stochastic opponent models and a closed-form likelihood for each trajectory §4.
4. Adversarial driving: planning high-risk manoeuvres and calculating the risk of future actions §4.
5. Real-time vehicle dynamics modelling, including learned models that are faster than tyre physics §4.
6. Balancing safety and performance in a single software stack §4.
7. A shared rulebook for autonomous racing §4.
8. Whole-stack latency, synchronisation and CPU/GPU scheduling §4.
9. Hardware built for high speed and shock resistance, instead of consumer hardware §4.

Hyperparameters

Not applicable: the survey proposes no method and runs no experiments, so it states no training or tuning values. Its one parameter table describes the platforms tab. 6.

Engine and battery data from tab. 6: F1TENTH 230 W motor and 55 Wh battery; Roborace 270 kW motor and 40 kWh battery; IAC 335 kW combustion engine. The survey gives no tyre, controller or planner parameters.

Results

Author-reported: these are numbers the survey gives for work it reviews. Each row names the setup as the survey describes it. Nothing here was measured by the wiki or re-run.

The survey gives no overall ranking of methods and reports no common benchmark. Its only aggregate number is the 233 covered papers, and it reports a rise in publications over the last four years §5 fig. 12.

Paper vs code

No official code. The survey has no implementation of its own. It links a GitHub repository with a structured list of the surveyed papers §5. That is a bibliography, not code, so no code field is set. Searches: the arXiv metadata has no comments or links; the LaTeX source links only that bibliography and an ICRA autonomous-racing workshop page (a footnote, not code); a GitHub repository search for the title words returned no results.

Pitfalls

• Scope ends at 2021. The survey covers work up to the end of 2021 §1.2.2 §5. The 2022 IAC races and later work are not in it. The IAC race dates are given as October 2021 and January 2022, with the text saying the competition "just finished" §3.2.
• Year conflict. The Williams et al. MPPI paper is dated 2016 in the planning table tab. 2 and 2015 in the control table tab. 3.
• Category count conflict. Section 2.3 says the control papers are "split into six subsections" §2.3, but it names four control categories in the text, and tab. 3 uses five labels (classic, MPC, optimisation, learning-based, drifting).
• Vague evaluation claims. The hardware section says "almost all" papers evaluate in simulation and "about half" also on real hardware §3. No counts or table support those shares.
• Boundary cases. Papers that sit between planning and control are assigned to one category by their focus §2. Search by method, not by section.
• Cited numbers come from other papers. The values in Results are the survey's summary of other works. Check the original paper before relying on one; the survey gives no conditions beyond those in the table.
• Missing ingredients for an implementer. No equations, parameter values or pseudocode are given for any surveyed method. Use §2.3 for the MPC loop shape only.
• Name inconsistencies in the bibliography. The planning table writes "Graghin" for Braghin and "Loiacona" for Loiacono tab. 2.

Lineage

• Builds on: the perception-planning-control pipeline of Pendleton et al., which the survey uses for its categorisation §1.2.2.
• Covers: the racing work of the cited papers, with the racing series and platforms in §3.
• Compared against: not applicable (survey).
• Companion resource: the authors' GitHub bibliography of the surveyed papers §5.
• Superseded by: not known from content in this wiki.