Comma.AI 开源其首次成功无人驾驶的旅行数据

来源：互联网发布：淘宝网店的推广方法编辑：程序博客网时间：2024/05/15 16:24

Comma.ai, the startup that George Hotz (aka Geohotz) founded to show that making driverless vehicles could done relatively cheaply using off-the-shelf components and existing vehicles, hasopen-sourced a dataset of 7.25 hours of highway driving.

It might not seem like a lot, but in terms of comparative datasets for highway driving out there, it is. And it’s what Hotz used to build the initial successful self-driving demo used toferry Bloomberg around for comma.ai’s big public debut.

“When I started this project, I didn’t want to have to put things in cars – I just wanted to play with the machine learning,” explained Hotz in an interview. “But I looked around and there was no good source of data to do that.”

Hotz points to the KITTI dataset and the more extensive and recent Oxford RobotCar dataset as a couple of sources, but these involved urban driving tests. Hotz was after a highway driving data source.

“There is not a good highway dataset to replicate what we had in Bloomberg and what we had in Nvidia, and I think it’s time that the whole world should be able to do this,” Hotz explained.

This doesn’t mean just anyone can take comma.ai’s data and turn their 1998 Ford Tempo into a self-driving superstar – but it’s a starting point. Comma.ai did not open source what they’re using to drive their test car, and the dataset represents where the company was at in terms of total data as of around six months ago; the company has obviously gathered more in the intervening time.

“I believe in being as open as possible without killing the host organism; we make sure we keep the company alive to open source more stuff,” Hotz told me about comma.ai’s approach to making their sets available to the general public. “There were a lot of missteps along the way to get this sort of data – no one needs to repeat them.”

Hotz emphasized that what comma.ai wants to do by open-sourcing data sets like this one is to enable the hobbyist community to accomplish more without having to do fairly basic, but time-consuming and resource-intensive work of collecting basic driving data for use in training machine learning systems. He points to DeepDrive, a self-driving car system which uses neural nets to drive virtual cars in Grand Theft Auto V as a prime example of the kinds of people they’re looking to help.

Helping hobbyists means building a talent pipeline, too: comma.ai can watch what people are doing with the data they open source, and bring in-house the brightest lights among that community.

“The other reason that we release things like this open source is that we’re incredibly confident that we have figured out what we’re going to ship, and our path to winning the positive feedback loop that is self-driving cars is there. And let’s pull some people up behind us, too.”

Ultimately, Hotz says, the comma.ai philosophy is simple – tell people exactly what they’re doing, but remain confident in the startup’s ability to do it faster, smarter and cheaper.

Comma.ai是一家致力于研发基于人工智能技术的汽车无人驾驶系统的公司，最近开源了一个7.25小时的公路驾驶的数据集。Comma.ai是著名黑客George Hotz（geohot）于2015年9月创办，并在2016年3月完成了未披露金额的种子轮融资的，同时geohot也是当年全球第一个破解iPhone 运营商锁定的年轻黑客。

Comma.ai是利用了CNN（Convolutional Neural Network, 卷积神经网络）快速破解汽车的网络，简化电脑对汽车的控制的——而且几乎只用了CNN。他们的车从一个摄像头把图像导入网络，从网络中输出命令，调整方向盘和速度，让车子保持在车道内。这样，系统内的传统代码很少，只有神经网络和一些控制逻辑。

虽然开源的数据看起来可能不是很多，但就公路行驶的数据集来说，它是很大的一个数。这是 Hotz 用来构建首次成功的自驾车demo，存放在Bloomberg，作为comma.ai’s的首次大公开亮相。

“当我开始这个项目的时候，我不希望汽车上有放东西—— 我只是想玩一玩机器学习。”在接受采访时Hotz解释道。“但我环顾四周，没有良好的数据来源做这件事。”

Hotz指出KITTI数据集和更广泛的牛津最近的RobotCar数据集是一对数据来源，但这些涉及城市的驾驶考试。Hotz是公路行驶数据源之后。

“没有一个很好的公路数据可以复制，所以我们开源在Bloomberg以及Nvidia上，我认为整个世界应该是能够做到这一点的时候了。”Hotz解释。

这并不意味着任何人都可以用Comma.ai的数据来把他们1998年的Ford Tempo变成自驾车巨星——但它是一个起点。Comma.ai没有开源他们正使用在测试车上的，该开源数据集代表的总数据是所在公司为大约半年前的，在这段时间内该公司已明显聚集了更多。

详情：Comma.ai open-sources the data it used for its first successful driverless trips

Comma.AI 开源其首次成功无人驾驶的旅行数据

相关链接