DiDi routes driverless R2 rides through the app that already matches its human drivers

Start with the computer, because it explains the partner. R2 carries 33 sensors, DiDi's full-stack Level 4 software and one central computing platform that combines three vehicle domains, with multilayer hardware and software redundancy ^[5][6]. Collapsing three domains into a single box means touching the vehicle's electrical architecture along with its power and network budgets, which is a vehicle program decision rather than something you bolt onto a car you bought finished. That is the practical reason R2 was jointly developed with GAC Aion, the electric-vehicle arm of Guangzhou Automobile Group ^[4]. GAC is also on the cap table: GAC Group led DiDi Autonomous Driving's $298 million Series C in October 2024, with DiDi participating ^[21].

The second half of the integration argument is dispatch. Riders hail an R2 from the same app that already handles human-driven trips, and DiDi says it can point its existing dispatch, mapping and fleet-management systems at the autonomous fleet ^[9]. Mechanically, that turns a geofenced Level 4 fleet into a supply pool inside a matching problem the company already runs: requests whose pickup and drop-off both fall inside the operating area can be offered to an R2, and everything else stays with drivers, so no rider sees a failed match. DiDi Global's 2025 annual report goes further, saying the program uses traffic data from the shared-mobility network for localization, prediction and vehicle control ^[11], with commercial viability framed as arriving through integration with an existing mobility platform ^[10]. Telemetry from human-driven trips is genuinely useful for map freshness and behavior priors, though it says nothing about how often the planner needs help, and the announcement publishes no autonomous mileage ^[8].

Which brings up the one safety figure DiDi did publish. More than 2,200 consecutive days of safe autonomous operations through its Huiju Port operations center ^[20] works out to 6.02 years ^[22]. Taking January 2026, when R2 was delivered ^[14], as the earliest possible date for the announcement ^[25], the streak reaches back to within about five months of the unit becoming an independent DiDi business in August 2019 ^[23][2]. So the metric sits close to "nothing we count as an incident has happened since we started," with no miles, rides or interventions underneath it ^[20]. For that number to mean anything in a reader's risk model, DiDi would have to define the event that breaks the streak and publish a denominator. The five-star C-NCAP and Euro NCAP claim has the same shape, asserting the standard without citing independent test reports or detailed ratings ^[19].

The cabin work is the part I would copy. Voice control for verifying the trip PIN, starting the ride, air conditioning, windows and ambient lighting ^[16], plus planned horn and voice prompts to help a passenger find the vehicle ^[18], are answers to the specific failure modes an absent driver creates. Pre-ventilating the cabin before pickup ^[17] is the least glamorous line in the release and probably the one that saves the most support tickets.

My read: for an operator that already owns the demand, this stack is the right tradeoff, because thin autonomous supply can hide behind human supply while the operating area grows. Scoring it requires the six categories of operating data DiDi withheld ^[24][8], and the company's own filing still says the technology had not been commercialized at scale as of the end of 2025 ^[12].