Li Auto i9 Assisted-Driving Demo Shows Stop-Sign Compliance Gap

Under China's road traffic sign rules, an octagonal stop sign requires a vehicle to bring its speed to zero at the stop line or before the intersection and proceed only after checking for safety. The i9 slowed in advance and crossed after confirming safety, but it did not execute the required full stop. Ifanr reported that the demonstration was therefore, strictly speaking, a case of assisted-driving rule violation.

The behavior reflects an algorithmic tradeoff, ifanr said. On Chinese roads, many drivers treat a stop sign as a yield sign, slowing and rolling through when the view is clear and no cross traffic is present; vehicles that fully stop can face honking or even rear-end collision risk. Many domestic assisted-driving systems have followed that customary pattern, the report said.

End-to-end models learn from real driving video clips and user feedback, the report noted, so they can acquire both experienced drivers' anticipation and some of their bad habits. The Li Auto demonstration, with its slow roll through the intersection, was described as the combined result of that learning and planning-control requirements, with traffic efficiency taking priority over regulation.

The issue is not unique to China. Ifanr cited research data showing that only 20.2% of drivers come to a complete stop at stop-sign intersections, a pattern also common among U.S. drivers. Tesla's FSD once allowed vehicles to roll slowly through intersections when no relevant vehicles or pedestrians were present. In 2022, the U.S. National Highway Traffic Safety Administration intervened and held multiple talks with Tesla, concluding that illegal behavior could not be packaged as a feature; Tesla then issued an OTA recall to disable the function.

Assisted driving can also be pushed into violations by road conditions that diverge from posted rules. Ifanr gave the example of a road with a 40 km/h speed limit where surrounding traffic commonly travels at 60 or 80 km/h. A system that strictly follows 40 km/h would not necessarily be safer, because large speed differences are a basic cause of rear-end collisions under traffic-flow theory. Human drivers adjust to the pace of traffic, and most assisted-driving systems allow a positive speed offset so they can keep up.

Roundabouts and unsignalized zebra crossings create similar pressure, the report said. Although rules require entering traffic to yield to exiting traffic and outer lanes to yield to inner lanes in roundabouts, real traffic often does not follow that order. At busy zebra crossings without pedestrian signals, some leading assisted-driving systems imitate experienced drivers by creeping forward to signal their intention and find a gap through continuous human-vehicle negotiation.

If a system adheres strictly to the letter of traffic law, ifanr said, it can quickly be rejected by users as inefficient and not intelligent enough. After Tesla FSD was tested in small batches in China, complaints appeared that it waited too long for pedestrians and did not understand domestic driving habits; a complete stop at a stop sign was cited by some as evidence that it was not localized enough. The report described a user contradiction: people say they want safety and legality, but once in the car they also want the system to follow traffic flow and seize gaps rather than behave like a newly licensed driver.

A new standard is set to change the calculus. GB 47955—2026, Safety Requirements for Combined Driving Assistance Systems of Intelligent Connected Vehicles, will take effect on January 1, 2027. It requires that the default cruise target speed of an L2 assisted-driving system not exceed the road speed limit recognized by the system. If the system detects a lower speed limit, it must lower the target speed automatically or after driver confirmation.

That means when a road is marked 40 km/h but traffic generally moves at 60 km/h, assisted driving cannot actively choose a compromise speed closer to the flow, according to the report. Speed limits that were not updated after road widening and temporary signs left after construction are all points the system must handle as deceleration requirements.

Ifanr argued that infrastructure and driver behavior also need to change as assisted driving becomes a normal part of traffic. Speed limits, road signs and markings should be adjusted, and human drivers need to rebuild stable right-of-way awareness, stopping when a stop is required and yielding when yielding is required. The report said it would be contradictory to demand strict compliance from assisted driving while treating a fully stopped lead vehicle as slow, because the more law-abiding vehicle could become the odd one out.

Adjusting algorithms for absolute compliance is technically relatively easy, ifanr said, but getting all road users to abandon long-standing habits and cleaning up unreasonable rules across the road network is a far larger governance task.