LESS GUESSWORK.
MORE TRACK TIME.

SARL is a virtual autonomous racing competition organized by Data Sorcerers. Student teams build the system that controls a virtual car’s steering, throttle and brakes. The mission is to discover and develop emerging Southeast Asian talent in automation, AI and control engineering.

SARL 2026 is open to active D4/applied bachelor’s, S1/bachelor’s and S2/master’s students, or their local equivalents, enrolled at an accredited higher-education institution in Southeast Asia. Teams may include students from different institutions and study programs. Bootcamps and similar non-degree providers are not eligible institutions. See Rules & Handbook for the full criteria.

Real-world racing experience is not needed to explore the concept. Programming fundamentals will help you build and debug a controller. Spellcraft focuses on engineered logic and control; Neuralcraft requires a meaningful learned neural component. The Developer Hub currently offers an educational controller preview.

The racing is virtual and autonomous. Your program controls the simulated car; manual driving, remote intervention or a human making race-time control decisions is outside the proposed format. You do not need to bring a physical racing car.

The current handbook and technical pages are labeled planning drafts. Final limits, eligibility and submission requirements will be published as approved documents. This FAQ is a guide to those materials; if it differs from an approved rulebook, follow the approved rulebook and its official clarifications.

Yes. The October 2026 roadshow and pre-event program will introduce SARL through seminars, webinars, campus roadshows and info sessions. Dates, speakers, campuses and joining links will appear on the Events page once announced.

Registration is planned for early November 2026; the exact opening date is TBA. Qualification runs across four online race weeks, with dates to be announced. Entries are not being accepted yet. Check registration status and Race Centre for updates.

No. Registration, online race weeks and the in-person final are free. The organizer will never request an entry or competition fee at any stage. If anyone asks for payment in SARL’s name, do not pay and report it to [email protected] or [email protected].

No. Every entry must be a team of at least two and no more than four eligible students. Team members do not need to attend the same institution or study the same major. Each participant may join one team in one division.

The proposal allows each person to join one team and each team to enter one division. Prepare one entry under either Spellcraft or Neuralcraft. Check the final handbook for the confirmed rule before registering.

Prepare your division, team name, captain, two to four participant names and email addresses, country, city, accredited institution, degree level and study program. The live process may request proof of active enrollment and institutional accreditation.

No. Entries are not open yet, and the website does not currently reserve competition slots or send participant confirmation emails. You will need to complete the live registration process after the organizer opens entries.

Registration is not currently open and no confirmation emails are being sent. For a live submission, keep the receipt or entry ID, check your spam folder and follow the published support instructions before submitting again. The live confirmation-email process will be announced when registration opens.

The participant change policy and its deadline have not been published. Do not create another entry to work around a correction. When live registration opens, keep your entry ID and follow the organizer’s published correction process; changes are not guaranteed.

Students from accredited public or private universities, institutes, colleges and polytechnics in Southeast Asia may enter. Bootcamps and similar non-degree providers do not qualify as the participant’s institution. All study programs are welcome, provided every team can develop and explain its code.

Yes. Team members may come from different eligible institutions, countries and study programs. Every member must independently satisfy the active-student and degree-level requirements.

Spellcraft is the Logic Drivers Track: you engineer decisions using rules, equations, optimization and classical control. Neuralcraft is the Neural Drivers Track: a trained neural component must meaningfully influence driving. Both use equivalent vehicles and evaluation scenarios. Compare the methods in Track Details.

Examples include PID, Pure Pursuit, Stanley control, Model Predictive Control, state machines and handwritten fuzzy rules. The proposed boundary is that no trained model participates in the runtime driving decision path. See the Spellcraft guidance for examples and deliverables.

Examples include deep reinforcement learning, imitation learning, neuroevolution, neural trajectory generation and hybrid neural control. A neural model may return actions or supply a target that another controller executes, but it must meaningfully affect driving. Review the Neuralcraft guidance.

The algorithm’s name does not decide the division. Offline genetic search that only tunes a classical controller can fit Spellcraft. Neuroevolution that produces neural weights used for driving can fit Neuralcraft. Describe what runs inside the driver at race time.

AI-assisted development outside the in-person final must be disclosed and remain within the published rules. During the 18–20 December final, teams may not use an LLM or generative coding assistant to build the driver, and an LLM may not run as part of the submitted system. Code determined to have been generated with prohibited assistance during the final receives a DNF.

A hybrid system can fit Neuralcraft when the trained neural component meaningfully affects the vehicle’s actions. For example, a neural planner could supply a target trajectory to a classical controller. The proposed deliverables include evidence explaining the neural component’s contribution.

No. Under the proposed boundary, tabular reinforcement learning without a neural component does not meet Neuralcraft’s neural requirement. A trained policy disguised as a lookup table also does not become a classical Spellcraft controller. Ask for an official category clarification if your method sits outside the published examples.

The website currently provides a controller contract preview and an educational example. These are not the official simulator or final SDK. Installation instructions, supported environments, release dates and versions will be announced in the Developer Hub and Development Updates.

Final development-device requirements will follow the simulator release. Teams are responsible for their development and training resources; organizer-provided GPU training is not promised. The draft evaluation environment proposes CPU-only inference, so do not assume a GPU will be available during official runs.

The current educational example uses native Node.js. That does not establish the final supported-language list. The official SDK and packaging guide will define supported runtimes, communication methods and dependencies. Prepare around the observation–action concept rather than assuming the example is the final API.

The preview describes vehicle and road observations as input, with steering, throttle and brake commands as output. It also illustrates readiness and reset operations. Read the controller contract preview for example units and ranges; final field definitions will come with the SDK.

The proposed championship uses offline inference with dependencies and weights packaged in the submission. External commercial inference APIs are outside that format. Check the final network restrictions before designing a driver that relies on online services.

The planning draft proposes 4 vCPU, 8 GB RAM, no GPU, a 20 Hz action rate and a 100 ms action deadline. These are not confirmed hardware entitlements or final limits. Use the versioned technical handbook as the reference once approved.

The proposal includes packaged runtime code, locked dependencies, configuration, a method report and an integrity declaration. Neuralcraft also includes frozen weights and a model card describing architecture, training, data and compute. Read shared submission guidance; the final packaging instructions and deadline are TBA.

Disclose reused code, weights, data, licences and AI tools. For the final, the requirement is to build a new AI driver from scratch on 18 December. Permitted libraries, pretrained components and reuse boundaries are not yet defined; do not assume an existing trained driver can be carried into the final.

No. Race Preview, the Spellcraft versus Neuralcraft illustration and generated city scenes explain the concept. They are not the official simulator, an approved circuit layout or confirmation of a physical street race. Only published simulator releases and official scenario packages define competition conditions.

Enter as a team of two to four eligible students. Each of four online race weeks accepts up to 50 teams and advances its top five, producing 20 finalists. In Central Java, finalists build from scratch on 18 December, qualify and evaluate on 19 December, then run the Grand Prix on 20 December 2026. See the full format.

The winner is the finalist with the fastest valid recorded Grand Prix race time. Qualifying times on 19 December determine the starting grid for 20 December. Lap count, timing, penalties and tie-breaks will be detailed in the sporting regulations.

A DNF means “did not finish.” Proposed causes include exceeding the lap-time limit, illegal checkpoint progression, overturning, prolonged full-footprint off-track travel, lack of progress and repeated action failures. The exact thresholds are in the DNF draft and will be finalized in the evaluator.

A time penalty adds to a run’s adjusted time; a DNF ends the attempt without a valid finish. The draft includes penalties for minor off-track episodes, collisions and late or invalid actions. Multiple applicable penalties may accumulate. Read the sporting rules rather than assuming every mistake ends the run.

Finalists will build their AI from scratch on 18 December for a new track determined by the organizer. The release procedure, permitted starting materials and practice access will be specified in the handbook.

Participants can develop and tune during the designated build and evaluation periods on 18–19 December. The draft still requires frozen model weights during an official run. The final submission deadline and reset contract will be published before the final.

The draft allows a written objection with a run ID, incident time and supporting reasons, within a proposed 20-minute review window. Verified organizer infrastructure failure may justify a rerun; a failure in participant code does not automatically do so. The official channel and final procedure will be published in Rules & Handbook.

Each online race week accepts a maximum of 50 entries. The top five from each of four weeks advance, making 20 finalists. Week assignment, repeat-entry rules and division allocation are TBA.

The four race weeks are online. The 20-entry final is held in person in Central Java, Indonesia, on 18–20 December 2026. The exact venue and attendance arrangements will be announced.

Travel, accommodation and personal expenses are the responsibility of each finalist team. During the scheduled in-person final program, the organizer provides lunch, dinner, snacks and mineral water at no charge. Final service times and dietary-information arrangements will be announced with the venue guide.

Start with this FAQ, the relevant Resources page and the latest Race Centre announcement. For participant questions or suspected payment requests, email [email protected] or [email protected]. Never send passwords, access tokens or unnecessary personal data.

No. The final begins with a fresh build on 18 December for a new organizer-selected track. Pre-built models and existing driver implementations are prohibited. Teams may use printed notes, printed documentation and printed code references, but may not use an LLM or generative coding assistant during the final. The final handbook will define the permitted libraries and starter materials.

Teams may use printed notes, printed documentation and printed code references during the offline final. Pre-built models, existing driver implementations, LLMs and generative coding assistants are prohibited. Organizer-provided starter materials and approved offline documentation will be defined in the final handbook.

Yes. Submissions may be reviewed together with required source history and organizer-provided development records to verify compliance with the from-scratch build and no-LLM rules. Participants will receive the final collection, retention and review terms before the final.