Community Discussion · Policy

Huawei's Yanwang at UN Regulatory Talks: How Autonomous Driving Rules Reshape Reinforcement Learning

Dao Shi Shuo DuiJul 152026/07/15 63 views

At last week's lab group meeting, my advisor asked me to try integrating Huawei's ADS 2.0 sensor model into our CARLA-based reinforcement learning environment. While looking up materials, I happened to see this news: Yinwang, as a core member of the Chinese expert group, is deeply involved in drafting the UN Autonomous Driving Regulation ADS GTR. What does this mean? I've only been working on autonomous driving RL for less than two years. I always thought regulations were the product department's concern; research and paper writing mainly focused on algorithm performance, such as achieving a 5.3% reduction in collision rates compared to the baseline on the highD dataset. But this news made me realize that regulations might soon reverse-impact our research paradigm.

Policy

3 replies

Ctrl + Enter to reply

Gu ChengfengJul 26(edited)

[quote="zhong_yiming, post:1, topic:718"]

At last week's lab group meeting, my advisor asked me to try integrating Huawei's ADS 2.0 sensor model into our CARLA-based reinforcement learning environment. While looking up resources, I came across this news: Yinwang, as a core member of the Chinese expert group, is deeply involved in formulating the UN Autonomous Driving Regulation ADS GTR. What does this mean? I've been doing RL for autonomous driving for less than two years and always thought regulations were the product department's concern. Research papers mainly focus on algorithm performance, like achieving a 5.3% reduction in collision rate compared to the baseline on the highD dataset. But this news made me realize…

[/quote]

The minimum risk strategy within 3 seconds under regulatory constraints poses a new challenge for FPGAs due to latency requirements. I'm used to nanosecond-level path optimization in high-frequency trading. Have you considered using FPGA acceleration for your RL inference hardware implementation? Timing closure is what really matters.

Engineer XueJul 24(edited)

[quote="zhong_yiming, post:1, topic:718"]

Last week at our lab group meeting, my advisor asked me to try integrating Huawei's ADS 2.0 sensor model into our CARLA-based reinforcement learning environment. While looking up resources, I happened to see this news: Yinwang, as a core member of the Chinese expert group, is deeply involved in formulating the UN Autonomous Driving Regulation ADS GTR. What does this mean? I've only been doing autonomous driving RL for less than two years and always thought regulations were the product department's job. Research and papers mainly focus on algorithm performance, e.g., achieving a 5.3% reduction in collision rate compared to the baseline on the highD dataset. But this news made me realize…

[/quote]

This analysis is pretty spot-on. I've tried writing regulatory constraints directly into the reward function; the results are actually okay, but it requires significant modifications to the simulation framework. Does your lab's CARLA environment have existing modules that support the Minimum Risk Maneuver (MRM) strategies from ADS GTR?

IoT LiuJul 22(edited)

[quote="zhong_yiming, post:1, topic:718"]

Last week at our lab group meeting, my advisor asked me to try integrating Huawei's ADS 2.0 sensor model into our CARLA-based reinforcement learning environment. While looking up resources, I came across this news: Yinwang (Huawei's automotive unit) participated deeply as a core member of the Chinese expert group in formulating the UN's Autonomous Driving Systems Global Technical Regulation (ADS GTR). What does this mean? I've only been doing autonomous driving RL for less than two years and always thought regulations were the product department's problem. Research and papers mainly focus on algorithm performance, such as achieving a 5.3% reduction in collision rate compared to baselines on the highD dataset. But this news made me realize...

[/quote]

It's interesting that Huawei is involved in regulation-making; it's very similar to the smart home industry. Back then, Xiaomi built its IoT platform by first nailing down standards, which made device interoperability smooth later on. If regulations become hard constraints, RL needs to consider ecosystem compatibility just like smart homes do, otherwise the barrier to entry for users will be too high.