Self-Driving Cars in 2025: What's Actually Working and What Isn't

Self-Driving Cars in 2025: What's Actually Working and What Isn't

I took a Waymo One ride in San Francisco last month. No driver. Nobody in the front seat. The car navigated complex urban intersections, handled a construction zone with flaggers, merged onto the freeway, and delivered me to a downtown restaurant. The experience was unremarkable in the best possible way — it just worked.

That unremarkable competence is the story of autonomous vehicles in 2025. The technology that was going to transform transportation "in five years" from 2015 onward has arrived in a much more limited, geofenced form than the early hype suggested. But within its current boundaries, it works reliably.

Understanding what's actually happening requires separating the technology's genuine progress from the narrative that's surrounded it — including Tesla's marketing language, Waymo's carefully managed operations, and the broader industry that went through a hype-crash-and-partial-recovery cycle over the past decade.

The SAE Autonomy Levels: What They Actually Mean

Self-driving discussions are only meaningful with a shared framework. The SAE defines six levels:

• Level 0: No automation. Human does everything.
• Level 1: Single-function assistance (adaptive cruise control, lane centering).
• Level 2: Multiple simultaneous functions (adaptive cruise + lane keeping). Human must remain attentive.
• Level 3: Conditional automation — system drives in specific conditions; human can disengage but must be available to retake control.
• Level 4: High automation — system can complete trips without human intervention in defined operational domains (specific geographic areas, weather conditions).
• Level 5: Full automation — any road, any conditions, no human required.

Where things actually stand in 2025: Level 2 is mass market (available in many new vehicles). Level 3 is emerging (Mercedes DRIVE PILOT available in Nevada and California in specific conditions). Level 4 is operational in geofenced areas (Waymo, Baidu Apollo, Cruise before suspension). Level 5 does not exist in any commercial product.

What Waymo Has Actually Built

Waymo is the unambiguous technological leader in deployed autonomous vehicles, and their operational data is the clearest evidence of how far the technology has progressed.

Commercial operations:

• San Francisco: Driverless paid rides expanded to most of SF proper
• Phoenix: Largest operational area, longest operational history
• Los Angeles: More recent expansion
• Austin and Atlanta: Announced expansion markets

Safety data: Waymo has published comparisons showing their vehicles have significantly lower injury-causing crash rates than human drivers in comparable conditions. This is meaningful data — not anecdote, not test conditions, but real commercial operations with real passengers and real conditions.

The limitations:

• Geofenced operations — Waymo doesn't operate everywhere. They map areas extensively before deploying, and their system depends on high-definition maps that may not be available in new areas.
• Weather limitations — heavy rain, snow, and fog reduce performance. Waymo's Phoenix operations benefit from consistently clear weather.
• Edge cases — unusual situations (extreme road damage, unexpected constructions, atypical human behavior) still cause hesitations and occasionally uncomfortable responses.

Why it matters: Waymo is demonstrating that Level 4 autonomous vehicles are commercially operable in real urban environments with real passengers. This is not a test program — it's a commercial service generating revenue and operating at scale.

Tesla FSD: The Reality vs. The Marketing

Tesla's "Full Self-Driving" capability is one of the most contentious marketing choices in the automotive industry, and understanding it requires separating capability from label.

What FSD actually does:

• Highway driving: Manages lane changes, following distance, and speed on freeways
• City driving: Handles traffic lights, stop signs, unprotected turns, and complex urban intersections
• Navigate on Autopilot: Complete routes from highway on-ramp to off-ramp

What FSD doesn't do:

• Operate without driver supervision (legally or in practice)
• Function in all weather conditions reliably
• Handle all edge cases without errors that require human intervention

The training approach: Tesla's camera-only approach leverages data from millions of Tesla vehicles. Each time a human intervenes in a situation FSD mishandles, that data trains the next version. At scale, this creates a powerful feedback loop.

The ongoing debate: Tesla has consistently argued that cameras plus AI are sufficient for full autonomy (the "vision-only" hypothesis). Most academic researchers and competitors argue LiDAR provides safety-critical redundancy that cameras cannot replace. This isn't resolved, and real-world safety data at scale will eventually answer it.

Regulatory scrutiny: The NHTSA has multiple ongoing investigations into Tesla Autopilot and FSD. Several countries have required Tesla to rename its product due to misleading implications.

The Competitors Reshaping the Landscape

Baidu Apollo: China's most deployed autonomous vehicle platform. Apollo Go robotaxis operating in multiple Chinese cities with thousands of rides. More aggressive government support for AV deployment in China than in the US.

Mobileye: Intel spinoff providing ADAS (Advanced Driver Assistance Systems) chips and now developing a robotaxi platform. Technology supplier to major OEMs; their SuperVision platform provides Level 2+ capabilities in millions of vehicles.

Cruise (GM): Suspended its driverless operations in October 2023 following a serious incident in San Francisco. Underwent significant leadership changes and operational review. Future unclear; GM has continued investing but with a much more conservative operational posture.

Zoox (Amazon): Developing a purpose-built autonomous vehicle (not a retrofitted existing car) for robotaxi deployment. Still in testing; the purpose-built approach may offer advantages over retrofitted vehicles.

Pony.ai: China-based autonomous vehicle company that went public in 2024. Operations in China and US.

The Regulatory Reality

Technology isn't the only barrier to autonomous vehicle deployment — regulation is equally determining.

United States: Patchwork of state regulations. California, Nevada, Arizona, and Texas have been most permissive for testing and limited commercial deployment. Federal regulatory frameworks from NHTSA are evolving but lag technology development.

European Union: More cautious regulatory approach. The EU has approved Level 3 systems (Mercedes DRIVE PILOT in Germany) and is developing frameworks for Level 4, but deployment is slower than the US and China.

China: Most aggressive government support for AV deployment. State backing for Baidu and other national champions. Less litigious environment and more willingness to accept some risk in pursuit of technological leadership.

Insurance: The current insurance framework was designed for human-operated vehicles. Determining liability when an autonomous vehicle causes an accident — is it the manufacturer, the software developer, the operator, the passenger? — is unresolved in most jurisdictions.

The Realistic Deployment Timeline

Now (2025): Level 4 robotaxis in select cities. Level 2/2+ in many new vehicles. Level 3 emerging in specific market segments.

2026–2028: Waymo-type services expand to 20–30 cities. Autonomous highway driving (L4) reaches commercial deployment for trucks in favorable conditions. Level 3 becomes more common in premium vehicles.

2028–2032: Autonomous trucking on major interstate corridors becomes significant. Robotaxi services reach hundreds of cities. Personally owned Level 4 vehicles begin entering the market (priced as premium).

2033+: Level 4 becomes mainstream in new vehicle sales. Most new vehicles capable of Level 4 highway autonomy. Full Level 5 remains rare or absent.

What Changes When Self-Driving Scales

The downstream effects of widespread autonomous vehicles are significant:

Trucking and logistics: 3.5 million truck drivers in the US. Autonomous highways could significantly reduce the need for long-haul drivers within a decade. The impact will be gradual but substantial.

Urban transportation: If reliable robotaxis reach $0.50–$1.00/mile (comparable to cost estimates from operational programs), car ownership in dense cities becomes economically irrational. This has enormous implications for urban planning and car manufacturing.

Insurance: An industry built on human error (most accidents are human-caused) faces fundamental questions when the primary cause of accidents changes.

Real estate: The parking space that currently accounts for 15–30% of urban land area becomes available for other uses as private vehicle ownership changes.

Frequently Asked Questions

Are self-driving cars available in 2025?

Level 4 robotaxis (no driver required) are commercially available in San Francisco, Phoenix, and Los Angeles through Waymo. Level 2/2+ driver assistance is available in many new vehicles. Fully autonomous vehicles you can buy for personal use are not yet available.

Is Tesla Full Self-Driving actually self-driving?

No. FSD is Level 2+ — it requires an attentive driver ready to intervene. It handles many driving tasks well but is not capable of unsupervised operation. The "Full Self-Driving" name has been widely criticized as misleading.

What is the difference between Waymo and Tesla's approach?

Waymo uses LiDAR + cameras + radar with extensive mapping; Tesla uses cameras only with AI trained on fleet data. Waymo has demonstrated higher reliability in current operations; Tesla's approach is more scalable if the vision-only hypothesis proves sufficient.

When will self-driving be mainstream?

Level 4 in controlled environments is available now and expanding. Mainstream personal vehicle Level 4 is likely 2030–2035. True Level 5 (anywhere, any conditions) may be 2035+.

Final Thoughts

Autonomous vehicles in 2025 are simultaneously more advanced than the skeptics claim and less advanced than the original hype promised. The Waymo commercial service is real, operational, and improving. Tesla FSD is genuinely capable but not self-driving in the way its name implies.

The 2030s will see substantial disruption to transportation, logistics, and urban planning as the technology matures and regulatory frameworks catch up. The question isn't whether autonomous vehicles will transform transportation — it's how long the transition takes and which industries get disrupted first.

For the broader AI technology trends that are reshaping industries alongside autonomous vehicles, the AI agents guide covers the software autonomy developments happening in parallel with physical vehicle autonomy.