惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

博客园 - 三生石上(FineUI控件)
月光博客
月光博客
人人都是产品经理
人人都是产品经理
Google DeepMind News
Google DeepMind News
M
MIT News - Artificial intelligence
Vercel News
Vercel News
MyScale Blog
MyScale Blog
爱范儿
爱范儿
博客园 - 司徒正美
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
IT之家
IT之家
H
Help Net Security
Last Week in AI
Last Week in AI
阮一峰的网络日志
阮一峰的网络日志
酷 壳 – CoolShell
酷 壳 – CoolShell
L
LangChain Blog
罗磊的独立博客
Stack Overflow Blog
Stack Overflow Blog
宝玉的分享
宝玉的分享
博客园 - 聂微东
云风的 BLOG
云风的 BLOG
J
Java Code Geeks
博客园 - 叶小钗
D
Docker

comma.ai blog

COMMA_HACK 7 Introducing chestnut Bugs that broke driving openpilot 0.11.1 Introducing commacare openpilot 0.11 COMMA_HACK 6 Owning a $5M data center openpilot 0.10.3 Introducing the comma four openpilot 0.10.1 openpilot 0.10 Learning to Drive from a World Model
openpilot 0.9.9
openpilot te · 2025-06-19 · via comma.ai blog

4 minute read

New driving model

As we prepare to ship driving models trained in our new ML Simulator, we are porting over small but necessary changes to the existing reprojective simulator.

We do not expect this release’s model to perform much differently than the previous release’s model, since this change is minor and does not alter the core components of the training setup.

Training at 5FPS

For compute efficiency, the ML Simulator produces frames at 5FPS rather than the 20FPS used at run-time.

The Filet-o-Fish model (#34798) is trained at 5FPS in the reprojective simulator to emulate this change and validate that the rest of the training stack (noise model, planner, etc.) does not regress.

Note that the model still runs at 20FPS at run-time, but its inputs are subsampled at 5FPS to match the training setup.

Open source reports

We found ourselves screenshotting these reports often to share them on Discord, so we decided to open-source them!

Here’s one example from analyse_cones showing the difference between the vision model outputs and the real trajectory driven by the human driver for a complex scenario involving cones.

There are lots of other tests; check out the full set of reports for the model in this release here, and look around for new models and new reports as we introduce more of them.

Online lateral delay estimator

lagd is a new daemon that learns your car’s lateral time delay online, which is the time it takes for the car to reach the desired lateral acceleration commanded by openpilot. This includes system processing latency, steering actuator delay, and steering response lag.

This lag varies across car models and even between individual cars of the same model, and both the driving model and the torque parameter learner rely on an accurate lag for precise compensation.

In this release, lagd runs in a passive shadow mode for validation, and if all goes well, the next release will enable the use of the learned value.

External USB GPU

Thanks to tinygrad’s USB GPU driver, you can now plug a full desktop GPU into the auxiliary USB-C port of your comma 3X (#35172).

This release doesn’t yet take advantage of the GPU, but in future releases, we’ll ship two classes of models: on-device and external GPU. There’s still room to grow for on-device models, but as we develop models whose performance scales with inference compute, we’re prepping the infrastructure to support external GPU-grade models.

This is a @comma_ai 3X with more compute power than a HW4 Tesla. It's a 9070XT + ADT-UT3G dock connected to the USB port.

Those are FP16 GEMM FLOPS, should be 389 TOPS of int8. Not a tech demo, shipping to openpilot soon. Excited to see how people mount a GPU in their car. pic.twitter.com/SBX6fm5mFV

— the tiny corp (@__tinygrad__) May 19, 2025

AGNOS 12

This is a minor OS update to fix display tearing and get back to 60FPS (agnos-kernel-sdm845#86). It also adds support for custom eSIM profile loading on comma 3X, thanks to @greatgitsby (agnos-builder#472)!

Cars

Tesla support

All Tesla Model 3 2019-25 and Model Y 2020-24 are now supported thanks to lukasloetkolben (opendbc#1263)!

Welcome to the fleet, Tesla 🫡pic.twitter.com/1E6E1GnA2p

— comma (@comma_ai) May 22, 2025

Because Tesla imposes high standards on the lateral and longitudinal APIs in their cars, control is very precise. openpilot interfaces with these same actuators, so the driving model is given unparalleled control in carrying out desired actions, whereas these APIs are slightly less refined on other cars.

With openpilot, you maintain both stock Automatic Emergency Braking and Lane Keep Assist (opendbc#2190). If your car is capable of Actually Smart Summon, that still works too (opendbc#1801).

For this port, we redesigned the lateral angle-control safety to support a consistent max lateral acceleration and jerk across all speeds (opendbc#2115). This translates to capable and smooth steering at all speeds. Our other angle-control cars will be moved over to this new angle limiting safety scheme soon, such as Toyota RAV4 2023+, Ford vehicles, and Nissan vehicles.

Bug fixes

  • Fix dash warnings on startup on some cars with comma three (#35287)
  • GM: fix intermittent dashcam alert on some older cars (opendbc#2033)
  • Hyundai: fix enable button press regression on some older cars (opendbc#2262)

Enhancements

Car Ports

  • Ford Escape 2023-24 and Ford Kuga 2024 support thanks to incognitojam! (opendbc#1922)
  • Hyundai Nexo 2021 support thanks to sunnyhaibin! (opendbc#933)
  • Tesla Model 3 and Y support thanks to lukasloetkolben! (opendbc#1263)
  • Lexus RC 2023 support thanks to nelsonjchen! (opendbc#2082)

openpilot is built by comma together with the openpilot community. Get involved in the community to push the driving experience forward.

comma is also hiring if you want to work on openpilot full-time.