The latest of our interviews with key companies and personalities in ADAS/AV and dual-use sensors, in the runup to our DVN Sensing & Applications Conference on 17 – 18 November in Stuttgart.
In this edition, we introduce an interesting hub from China, i4Research. An ADAS/AV vehicle test and benchmark entity located in Shanghai. I4Research Shanghai Ruisiqi Information Technology Service Co., Ltd. – Interview by Dr Jürgen Dickmann, Senior Advisor, DVN
DVN: Who and what is i4R?
I4R: i4Research is the world’s first dedicated intelligent driving data platform, founded by a team with deep automotive experience in Shanghai.
Our mission is to build an efficient tool platform that improves the autonomous driving experience. Our core value lies in rigorous public road testing across China’s most complex urban environments, delivering high-fidelity, real-world performance data.
For challenging scenarios, we also operate a professional third party test field in Nantong for advanced ADAS validation. Across more than 110 standardized test protocols, we have evaluated over 70 Chinese production models.
DVN: From your recent testing cycles in cities like Shanghai and Shenzhen, what are the top three ‘edge cases’ that still differentiate the leaders from the followers in the L2++ domain?
I4R: The gap between leaders and followers in L2++ systems is defined by how they handle scenarios that could be defined as standard driving. Based on our latest cycles, the three most differentiating edge cases are:
- Ride comfort under lateral disturbances – Followers often brake abruptly or hesitate excessively when a vehicle drifts near the lane boundary, causing passenger discomfort. Leaders apply smoother, predictive deceleration.
- Lane change efficiency in dense traffic – Followers take too long to complete a lane change, sometimes crossing a solid line or giving up and getting stuck in a jam. Leaders execute timely, confident merges even in slow moving congestion.
- Cut in detection from agile small vehicles – Sudden cut ins by two wheelers or tricycles remain a critical failure point. Followers frequently miss these fast, narrow obstacles, requiring driver takeover. Leaders handle them with earlier recognition and appropriate evasion or braking.
DVN: How do you transform raw, real world ‘disengagements’ into actionable insights for engineers validating their perception and planning algorithms for the Chinese market?
I4R: Because we capture multiple scenarios that trigger a takeover, our analysis goes far beyond simply identifying the failure. We record each test case in detail, and our trained drivers provide qualitative annotations – such as why a disengagement felt unsafe or inefficient. This extra layer helps engineers pinpoint deficiencies in their occupancy network or motion prediction.
This forms a ground truth of what the safe, expected behavior should have been, allowing teams to debug planning decisions in a structured, repeatable way.
DVN: How does i4research provide objective ‘ground truth’ when OTA updates change software stacks so rapidly? How do you validate whether a quick OTA fix truly resolves an edge case?
I4R: The speed of OTA updates in China is indeed a challenge. We sometimes test the same car three or four times in a single year due to continuous updates. We provide an objective “ground truth” layer precisely to validate these hyper reactive fixes.
Our structured methodology ensures that every new OTA is stress tested on the exact scenarios that triggered the original failure. Because we run identical test protocols, we can directly compare the previous and new versions across every metric.
We then deliver post validation reports that answer: did the update resolve the root cause or just the symptom? For example, did it reduce disengagements by genuinely improving perception, or simply by slowing down the car so much that the edge case no longer occurred? This ensures the fix is truly robust.
DVN: How does your standardized benchmarking methodology help global Tier 1s and OEMs quantify ‘China Speed’, and what does your data suggest about the current performance gap between local Chinese software stacks and international solutions?
I4R: Our standardized benchmarking methodology quantifies “China Speed” by tracking how quickly OEMs iterate capabilities on high frequency urban routes. We record every OTA and monitor market evolution week by week.
For most global Tier 1s and OEMs, the initial gap was in upgrading their vehicle electronic architectures, local leaders like Xpeng and NIO have already deployed next generation solutions. Today, the remaining gap lies in handling L2++ edge cases, especially navigating not only urban chaos but also complex highway interchanges.
That said, we expect this gap between international solutions and local players to narrow significantly over the next months.
DVN: How has your testing approach evolved to evaluate ‘black box’ end to end AI systems, and what metrics do you use to measure their reliability compared to traditional rule based stacks?
I4R: With end to end systems, we have shifted focus from pure quantitative KPIs to a richer mix of behavioral and subjective metrics. Our trained drivers now provide much more detailed subjective evaluations, because two systems can have identical objective performance (e.g., same number of disengagements) yet feel completely different to a customer.
To capture that, we have defined three categories of disengagement:
- Risk based (safety critical, immediate takeover required)
- Inefficiency based (system hesitates or takes an illogical maneuver)
- Rule violation (crosses a solid line, misses a traffic signal)
This framework helps us compare black box systems against rule based ones not just on safety, but on driving naturalness and user trust.
DVN: Do you see a difference in driving behavior depending on the sensor suite used?
I4R: Yes. While all high level systems share the same ambition, their sensor configuration directly defines their strengths and weaknesses on the road. A pure vision system may deliver average but consistent behavior, whereas a LiDAR equipped system can be extremely precise, sometimes at the cost of producing more abrupt, sudden maneuvers.
In contrast, cars with lower ADAS levels quickly reveal performance degradation at night or in poor weather, as their minimal sensor suites struggle with reduced visibility.
DVN: With your growing partnership with DVN and your presence at international conferences, how do you see the role of i4research expanding in the global safety validation ecosystem? What is the one thing a European OEM should look at in your data before finalizing their ADAS strategy for China?
I4R: Analysing the Chinese market is complex and time consuming. It is hyper competitive, and OEMs need objective engineering insights, not just PR claims. Before finalizing a strategy for China, European OEMs should examine our data on real world system interventions for L2++ edge cases.
This reveals the critical gap between a system that is merely acceptable in Europe and one that truly competes in Shanghai or Shenzhen traffic. Our database shows exactly where international stacks break down, not in theory, but on actual roads. That single insight can save days of misdirected development.
DVN: Dante, thank you for the insights into this interesting company, and your pariticpation as speaker at our 9th DVN Conference on 17/18 November in Stuttgart.
