Why might the driving performance of the same AD solution still vary when installed in different vehicle models?
A vehicle may corner smoothly, yet exhibit steering response deviations when the same solution is applied to another model; identical deceleration commands may also result in different braking sensations across different vehicles. This does not imply a change in algorithmic capability.
Ultimately, an AD system controls a real physical vehicle, not an idealized model in a simulation environment.
Changes in a vehicle's hardware configuration, dynamic characteristics, and control interfaces can all affect the system's actual performance. Therefore, when reusing an AD solution across different vehicle models, the real challenge is not merely whether the software can be deployed, but whether the original solution can adapt to the new vehicle conditions while maintaining the expected driving performance.
01. Why Does the Same Algorithm Perform Differently Across Vehicle Models?
An AD system must translate perception results into driving decisions, which are then executed via the vehicle's steering, braking, and propulsion systems. This process involves the coordination of perception, planning, control, and the vehicle's hardware. Condition changes in any link can affect the final outcome.
Consider the sensors first. Even if two vehicle models use the exact same camera model, their mounting height, pitch angle, and orientation may differ. The relationship between the sensor coordinate system and the vehicle coordinate system must be calibrated; otherwise, systematic deviations may occur in information such as target positions and road boundaries. Multi-sensor fusion also requires handling the spatial relationships and time synchronization among different sensors.
Next, consider the vehicle itself. Different models may vary in wheelbase, gross vehicle weight, tire characteristics, and steering system response. Even if the target trajectory provided by the planning module is identical, the actual vehicle motion may not be the same. For instance, if the system expects the vehicle to follow a curve, the control module must calculate steering commands based on the vehicle's state. If the actual response of the steering system deviates from the original model, the vehicle may fail to accurately follow the expected trajectory.
The same applies to longitudinal control. The powertrain response, brake build-up process, and vehicle mass of different models can all affect actual acceleration and deceleration performance. This means there is no simple one-to-one correspondence between the driving intent output by the algorithm and the final actions executed by the vehicle. After migrating the solution to a new vehicle model, it is necessary to recalibrate the sensor extrinsic parameters, identify vehicle dynamics parameters, adapt control parameters and strategies, and verify their applicability under the new vehicle conditions through closed-loop testing on actual vehicles.
02. What Exactly Needs to Be Adjusted for Cross-Model Adaptation?
For cross-model adaptation of an AD system, three main aspects must be considered: parameter calibration, system adaptation, and whole-vehicle validation. Parameter calibration primarily addresses whether the system's description of the vehicle and sensors is accurate, and whether the actuator characteristics and control parameters match. For example, sensor mounting parameters, relevant parameters in the vehicle dynamics model, and certain calibration values in steering and longitudinal control may all need to be recalibrated or identified based on the new vehicle model, and then confirmed during subsequent control adaptation and whole-vehicle validation.
System adaptation involves broader tasks. Different vehicle models may adopt different sensor layouts, computing platforms, and vehicle interfaces. The definitions of vehicle state signals, the transmission methods for control commands, and feedback mechanisms may also differ. The R&D team needs to confirm whether the original software architecture and interfaces can be directly reused, which modules require modification, and which functions need to be reintegrated.
Finally, there is whole-vehicle validation. Even if sensor calibration is correct and control interfaces are functioning normally, it does not guarantee that the vehicle will achieve expected performance under various real-world driving conditions. The R&D team must also conduct tests to evaluate the vehicle's trajectory tracking, steering response, acceleration and deceleration performance, and system behavior under different speeds and road conditions.
These three tasks are interrelated but cannot replace one another. Calibration addresses parameter accuracy, system adaptation ensures the software matches the vehicle, and whole-vehicle validation confirms whether the entire system meets the requirements on the target vehicle model.
Of course, switching to a new vehicle model does not mean redevelopment of all modules, nor does it require retraining the entire AD model. If the vehicle platform, sensor configuration, and control interfaces are relatively similar, existing software, models, and parameters can be reused to a large extent; the greater the differences, the more work is typically required for re-adaptation and validation.
03. Why Should Automakers Participate in the Joint Tuning of AD Solutions?
The importance of vehicle model adaptation is also evident in the collaboration between AD suppliers and OEMs. In September 2026, Shenlong Automotive Technology and Momenta announced a global strategic partnership. Public information from both parties mentioned that they will combine their capabilities in vehicle R&D and chassis tuning to adapt core parameters such as steering, braking, and powertrain response, and conduct joint development tailored to the characteristics of vehicle models under brands like Peugeot and Jeep.
What makes this case noteworthy is not the specific algorithm used in a particular solution, but the issues that OEMs and AD technology suppliers must address together. The driving decisions of an AD system are ultimately realized through the chassis and actuators of a specific vehicle model. The vehicle's dynamic characteristics and the brand's requirements for driving feel can both influence how the system is adapted.
For OEMs, AD functions must not only be operational but also coordinate with the vehicle's powertrain, steering, braking, and other control systems. For AD suppliers, the reusability of a solution cannot be judged solely by software deployment speed; it also depends on whether the solution can be adapted, validated, and delivered under different hardware and vehicle conditions.
This is also where the value of platform-based development lies: unifying reusable software architectures, toolchains, and interfaces as much as possible to reduce ground-up development for each vehicle model, while retaining necessary adaptation space for different models. The solutions publicly presented by Horizon Robotics also emphasize supporting the development, validation, and deployment for different vehicle models and partners through standardized computing platforms, toolchains, and sensor solutions.
However, platformization does not mean using the exact same parameters for all vehicle models, nor does it mean a single validation can cover all models. The more reusable components there are, the greater the opportunity to improve cross-model development efficiency; yet, identifying, adapting to, and validating vehicle differences remain indispensable steps in mass production development. Expanding from one vehicle model to more tests not only the AD algorithm itself but also the transferability of the entire technical solution.
A truly mature mass production solution must enable the reuse of core capabilities while allowing the system to correctly understand vehicle states and accurately execute driving commands across different vehicles, with actual performance confirmed through whole-vehicle validation. Reusing an AD solution is not about making different vehicle models identical, but about enabling the same set of technologies to adapt to the differences among various vehicle models.