Autonomous driving is the craft that converts what a vehicle senses into motion decisions on public roads. The work runs from perception through fusion to planning and control, and each stage hires a different person; the stages share a vocabulary, not a bench. The one requirement they do share is road evidence: what shipped, in which operating domain, with which measured limits.
Volume sets the floor for the hiring market. EU registrations grew 1.8% in 2025 with battery-electric share at 17.4%, ACEA reported in January 2026 . The bar keeps rising on top of it: Euro NCAP announced in November 2025 its largest rating overhaul since 2009, scoring safe driving, crash avoidance, crash protection and post-crash response as distinct stages from 2026 . Both pressures reward released-feature evidence over algorithmic novelty, and that is what makes the pool thin.
Hiring challenges in autonomous driving
Advanced driver-assistance systems (ADAS) hiring starts at the SAE level boundary
The six SAE levels separate driver support (Levels 0–2) from automated driving features (Levels 3–5), and the human role changes completely across that boundary . CVs routinely blur it. Data annotation, demo scripting and simulation tooling all appear under one title, so the brief must state the level, the operating domain and the release responsibility before sourcing begins. UN Regulation No. 157 made automated lane keeping the first binding regulatory step for an automated driving system, with approval provisions the Level 3 bench now measures itself against .
The 2026 NCAP protocols sharpen the same boundary from the customer side, scoring assistance and avoidance separately from crash protection . A Level 2 feature owner and a Level 3 item owner therefore answer different interview questions. A pipeline that does not ask the level first spends senior perception hours screening people two levels away from the seat.
Autonomous vehicle perception hires on quantified failure modes
Autonomous vehicle perception built from cameras, LiDAR and radar fails in specific, repeatable ways: glare, spray, snow-covered lane markings, small overlapping road users, sensor blockage. ISO 21448 exists for exactly this territory, providing the argument framework for safety of the intended functionality where situational awareness comes from complex sensors and processing algorithms . Candidates should describe the failure modes they characterized, on which datasets, with which metrics, and the performance envelope they signed.
The alternative is the architecture monologue: a candidate who can narrate a detection stack end to end but cannot state a single measured limit. That profile belongs in research. The release seat needs the numbers, because the residual-risk argument that follows the feature is built from them.
Computer vision for the road is a weather-corner discipline
Computer vision on public roads is mostly a problem of corners: low sun, spray from a passing truck, snow covering the markings the model trained on. The 2026 protocols push testing into exactly these conditions, with urban scenarios built around two-wheelers, cyclists and pedestrians . Strong candidates describe which corners their data covered and which it did not; weak ones describe the model.
Skipping this probe means late-found perception failures that no amount of planning logic recovers. A detector that works on curated datasets and fails in February is a validation discovery, and validation discoveries at prototype stage are the most expensive kind.
LiDAR and radar choices pin the sensor set before the CV
The sensor set is an architecture decision, and LiDAR and radar fill opposite slots in it. Radar holds its behaviour in precipitation and dust; LiDAR carries the geometry cameras only infer. A candidate's value is set-specific: which mix they validated, why that mix, what the compute and cost envelope forced out, and which weather corners the set left uncovered.
Two candidates who both say "multi-modal perception" may never have touched the same physical sensor, and the interview only discovers that when the questions turn to mounting, alignment and field-of-view budgeting. Compute and thermal envelopes further constrain the set, and a candidate who has never traded sensor resolution against cost per unit has designed a demonstrator, not a vehicle.
Sensor fusion owns the redundancy arithmetic
Sensor fusion engineers answer the question nobody poses at architecture review: what the stack does when one modality drops out at speed. Time alignment, cross-modality checks, degraded-mode fallbacks and arbitration between conflicting hypotheses are the actual job. Candidates must show the failure the redundancy caught and the failure it missed.
The probing question is a dropout scenario with a number attached: one modality fails at 130 km/h, what is the decision latency, what does the fallback assume about the remaining sensors. Owners answer with traces; spectators answer with block diagrams.
Path planning must answer the end of the map
Path planning looks like graph search until the map stops. Construction zones, temporary layouts and occlusion guarantee the high-definition map lags reality, and the planner has to produce a legal, safe trajectory anyway. The interview test is a concrete corner: which constraint was violated in the scenario, which fallback the candidate owned, what the recorded data showed afterwards.
The hard case is not the lane change; it is the junction where every path leads into an occluded region and the planner must commit before it can see. Candidates who have only planned inside curated maps answer in principle. Candidates who have shipped answer with a failure log, which is the only evidence that matters in this seat.
Motion planning inherits actuator limits at the tyre
Motion planning meets physics at the last meter. Comfort-against-safety trade-offs, degraded friction, loaded and towing states, and the chassis's actual response time bound every trajectory the planner may propose. Planners hired without that fluency produce paths the vehicle cannot follow; the discovery arrives during track testing at full daily cost.
Ask which constraints the planner respected, which comfort-against-safety trade the candidate owned, and how the plan degraded when perception confidence fell. This is the seam where the autonomy stack meets vehicle dynamics, and it is where a programme discovers whether the planner and the chassis team have been talking. Vehicle dynamics engineers define the envelope; the planner's job is to stay inside it while the world moves.
Autonomous vehicle control arguments end at the safety case
Whatever the stack learns, the feature ships through documents. Functional safety under ISO 26262 attaches lifecycle and work products to safety-related E/E systems in series production ; cybersecurity engineering adds risk management across concept, development, production, operation and decommissioning ; type-approval regulation audits cybersecurity and software-update management before market entry .
Owners answer in work products: which hazard analysis, which verification report, which assessor challenge, what the audit found. Assessment quality determines whether scarce engineering time turns into a release. The useful questions are bounded: which operating domain, which sensor set, which limit was measured, which work product carries the signature, and which audit was survived. Hiring a demonstrator into an owner seat yields a feature that works in the demo and cannot ship, discovered at the gate where the release waits. If shortlists keep collapsing at the technical screen, the missing step is an engineer-led autonomy assessment before interview, not a wider keyword net.
References
- New car registrations: +1.8% in 2025; battery-electric 17.4% market share — European Automobile Manufacturers' Association (ACEA). (accessed 2026-09-28)
- Euro NCAP announces 2026 protocol changes to tackle modern driving risks — Euro NCAP. (accessed 2026-09-28)
- J3016_202104: Taxonomy and Definitions for Terms Related to Driving Automation Systems — SAE International. (accessed 2026-09-28)
- UN Regulation No 157 – Uniform provisions concerning the approval of vehicles with regards to Automated Lane Keeping Systems [2021/389] — EUR-Lex, Publications Office of the European Union. (accessed 2026-09-28)
- ISO 21448:2022 — Road vehicles — Safety of the intended functionality — International Organization for Standardization (ISO). (accessed 2026-09-28)
- ISO 26262-1:2018 — Road vehicles — Functional safety — Part 1: Vocabulary — International Organization for Standardization (ISO). (accessed 2026-09-28)
- ISO/SAE 21434:2021 — Road vehicles — Cybersecurity engineering — International Organization for Standardization (ISO). (accessed 2026-09-28)
- UN Regulations on Cybersecurity and Software Updates to pave the way for mass roll-out of connected vehicles — United Nations Economic Commission for Europe (UNECE). (accessed 2026-09-28)