Level 5 Autonomous Driving
Level 5 means a car drives itself anywhere a human could, in any weather, without a driver. Level 4, as at Waymo, works only in defined zones with high-precision maps. The hard part is the long tail of rare situations, such as roadworks, police hand signals, snowfall and unexpected behaviour. Missing are perception that understands novel situations, solid statistical proof of safety, reliability in snow, rain and fog, a legal framework without area limits and lower sensor and computer cost.
Open in the interactive tree →The SAE levels run from 0 to 5; level 5 means the vehicle drives in all conditions a human could handle. Level 4 - as at Waymo - works in defined zones with high-precision maps. The problem is the 'long tail' of rare situations: roadworks, hand signals from police officers, snowfall, unexpected behaviour.
As of now
Waymo (about 4,000 vehicles in 14 US cities) and Baidu (over 22 million rides) drive level 4 in mapped urban areas; Waymo's own data to June 2026 show 95% fewer serious-injury crashes than human drivers over 271 million rider-only miles. Tesla sells 'Full Self-Driving (Supervised)' with a driver and ties large-scale unsupervised operation to FSD v15 (late 2026 or early 2027). Nobody drives level 5.
What is missing
- Perception that understands rare and novel situations safely instead of relying on maps
- Statistically solid proof of safety over hundreds of millions of kilometres, plus explainable behaviour
- Reliability in snow, heavy rain and fog and on unmapped roads
- A legal and liability framework that allows driverless cars without area limits
- Cost: sensors and compute must fall to the level of a private car
Becomes possible once solved
- Mobility for people without a licence (elderly, children, people with disabilities)
- Far fewer road deaths
- Driverless trucks across whole continents
- Cities without large parking areas
- Mobility as a service instead of private ownership
Open steps
- Rare and novel situations High AI leverageHandle events the car has never seen, with perception and planning that generalize beyond mapped areas.
- Foundation-model driving policies High AI leverageTrain driving models that generalize from fleet data to new cities and roads, with predictable gains from more data.
- Statistical safety proof Medium AI leverageShow with confidence that a driver is safer than humans using limited real miles plus simulation and fleet logs.
- Snow, fog and unmapped roads Medium AI leveragePerceive and drive when lane markings, sensors and maps fail.
- Cheap sensors and on-board compute Medium AI leverageCut sensor and compute cost to private-car levels while keeping safety.
Where today's AI could help
High AI leverage. The unsolved part is long-tail perception, planning and safety proof, where AI plus simulation is the main tool; law, cost and weather limits remain.
- Generate rare and dangerous scenarios in generative simulation to train and test the driver
- Train perception and planning models on fleet data for rare objects and bad weather
- Estimate safety statistically from simulation and fleet logs to support approvals
- Distill large models into cheap on-board chips to cut sensor and compute cost
Shown so far
- Waymo's own data through June 2026 report 95% fewer serious-injury crashes than human drivers over 271.3 million rider-only miles in its mapped areas (company data). source
- In February 2026 Waymo introduced a generative world model based on DeepMind's Genie 3 that simulates rare events such as tornadoes or an elephant on the road for training and testing (according to Waymo). source
Prerequisites
- Satellite navigation (GPS)1995
- Deep Learning2012
- Robotaxis in Daily Service2020–2026