这家中国公司的逆天科技，彻底搅乱无人驾驶业态

无人驾驶领域又要放大招了。一家名不见经传的初创公司正在研究一种新型的无人驾驶技术，据这家公司称，该技术能让汽车利用“常识”，在不可控的环境中像人类一样驾驶。

在无人驾驶领域，多数公司的研究重点是如何改进传感器、感知和控制技术。而iSee公司的CEO赵一彪表示，他的公司开发的机器人，将是业内第一个能够真正理解路面上发生了什么的机器人程序。

大约一年前，赵一彪与麻省理工学院的实验室伙伴克里斯·贝克和科技创业人黛比·于（音译）一同创办了这家公司。该公司也获得了麻省理工学院的风投机构The Engine的资助。

赵一彪上周在接受《财富》杂志的电话采访时表示：“我们知道，‘看到’不等于‘理解’。目前，汽车已经实现了‘看到’，但它们并不能真的理解究竟发生了什么，别人在想什么，以及别人的意图是什么。”

而iSee公司则为汽车程序提供了一种特殊算法，使汽车可以在开放环境中与人类协作。该系统有两个主要组成部分：深度学习和常识引擎。

系统的深度学习部分与Waymo和Uber的现有技术大同小异。所谓深度学习，是指如果你对某件事练习得足够多，你就能下意识地做到它。对于人类来说，正是快速的下意识思考，才使你在驾驶时能够同时完成多个操作。而对于无人驾驶技术，也正是这种深度学习，才使得汽车能始终保持在自己的车道里行驶。

然而当你遇到特殊路况，你就需要推理和有意识的思考了。比如在高速公路上会车、变换车道或通过十字路口时，你都要预测其他车辆的活动，有时甚至要跟其他司机“谈判”，有时也得考虑其他可能性，以求安全通过。

赵一彪表示：“作为一个人类司机，我们会有意识地思考那些可能发生的情形，这是因为我们的头脑中有一个‘常识引擎’，因此，哪怕遇见一些从未见过的情形，我们也同样有能力去应对它。”

而无人驾驶汽车的常识引擎，则同样可以让汽车基于常识和过往的经验来应对新情况。

赵一彪介绍道，这个独有的常识引擎能帮助汽车“真正理解现在发生了什么，并且预测接下来两秒钟里，他们可能会做什么。”这样一来，机器人就能“在需要与环境中的其他司机进行交流甚至谈判时，做出安全的、策略性的决策。”

赵一彪三人在大约一年前建立了这个算法，并制作了一个模拟器。然后他们想到，为什么不在一台真正的汽车上做实验呢？黛比·于大方地借出了自己的一台混动SUV作为实验用车。

回忆起2017年冬天他们在车库里工作的经历时，赵一彪笑道：“我们只花了两周时间，就让那辆车实现了自动驾驶。这是一次非常有趣的经历。”

第一次实验成功后，iSee的程序又经过了多次修改完善。这个团队不仅利用模拟器进行测试，也在美国的多个州通过有人驾驶汽车实测了系统在真实路况中的表现。

这种技术使机器人得以流畅地与人类协作。虽然该技术在无人驾驶领域之外也有应用潜力，不过赵一彪表示，目前iSee将主要专注于无人驾驶领域。

“我们认为，无人驾驶是一个新兴市场，所有人都在朝着这个目标努力。而且现在市场已经准备好了，客户也已经准备好了，缺的只是实现它的技术，所以我们的当务之急，是让这个‘杀手级应用’能够管用。未来我们也可以将它扩展到其他应用上。”

随着常识引擎在技术上取得成功，iSee公司也希望它能被广泛接受——至少不要像行业领袖Waymo那样遭到车主的嫌弃。据称，自从Waymo在亚利桑那州投入测试以来，已因其过于保守的驾驶风格遭到了很多人类司机的不满。

对此赵一彪表示：“我认为这正是这个领域的一个公开挑战。我觉得对于常识理解的核心部分，现在连Waymo和Uber这些公司也没有搞明白，而我们正是专注于这一块的，我认为它将是让无人驾驶系统真正适应真实路况的一项关键技术。”

不过，这个目标什么时候才能实现呢？赵一彪表示：“它已经发生了，它不是未来，而是现在。”（财富中文网）

译者：朴成奎

A little-known name in the world of autonomous driving is paving the way for a new type of self-driving car—one that can use “common sense,” as the company calls it, to navigate an uncontrolled environment.

While most companies developing self-driving cars are focused on improving sensors, perception, and control, iSee CEO Yibiao Zhao says his company is the first to work on creating a robot that can really understand what’s going on.

Zhao founded iSee just about a year ago along with Chris Baker, his lab partner at the Massachusetts Institute of Technology, and Debbie Yu, who has a history with tech startups. The three are supported by MIT’s venture capital firm, The Engine, in Cambridge, Mass.

“We know that seeing is not equivalent to understanding,” Zhao told Fortune in a phone call last week. “Currently cars can see, but they cannot really understand what’s really going on and what other people are thinking, and what are the other people’s intentions.”

With iSee, the cars’ programming has a special algorithm allowing it to collaborate with humans in an open environment. The system has two components: deep learning and the common sense engine.

Deep learning is something other companies like Waymo and Uber have already established; it’s the notion that if you practice something enough, you’ll be able to do it unconsciously. In humans, it’s the fast, subconscious thinking that allows you to multitask while driving. In self-driving cars, it’s the type of learning that lets a car remain within a lane.

When you get to an obstacle, however, you’ll need reasoning, or conscious thinking. When you merge on the highway, change lanes, or come to an intersection, you need to predict the actions of other cars, negotiate with them, and consider different possibilities in order to make a safe decision.

As a human driver, “we’ll consciously think about those types of possible parallel futures,” says Zhao. “That is enabled by our common sense engine in our mind, and that gives us the ability to handle some new scenario that we never encountered before.”

In a self-driving car, the common sense engine allows it to navigate new situations based on a handful of past experiences and general knowledge.

This component, unique to iSee, helps the car to “truly understand what is going on, and to predict what they might do in the next two seconds,” says Zhao. This lets the robot “make safe and strategic decisions when they need to interact or even negotiate with the other drivers in the environment,” he says.

Once Zhao, Baker, and Yu had this algorithm established and passed through a simulator about a year ago, they figured “why not” try it on a real car, says Zhao. Yu generously agreed to let her car, a hybrid SUV, serve as guinea pig.

“We spent just two weeks, and we made the car driving,” says Zhao, laughing as he recalls how cold it was working in the garage in the winter of 2017. “It was a very fun experience.”

Since the success of that first experiment, iSee has gone through multiple variations of programming. The team tests their system with both a simulation engine and manned cars driving in multiple states.

This kind of technology, allowing robots to work fluidly with humans, has potential outside the industry of autonomous cars, but Zhao says iSee is focused on self-driving cars for now.

“We believe the self-driving car is the emerging market. Everyone is working so hard towards it, and the market is ready, the customer is ready,” he says. “What is lacking is this enabling technology, so we want to make this killer application work first. In the future, we can extend it to other applications.”

With the success of the common sense engine, iSee hopes to become widely accepted—without the controversies that have surrounded industry leaders like Waymo, which is reportedly hated by its human neighbors in Arizona due to the cars’ overly-conservative driving.

“I think that’s the open challenge in the field,” says Zhao. “There’s one single piece—that is this core part of the common sense understanding—and I think even Waymo and Uber, those companies, haven’t figured it out yet. We are laser focusing on that and I think that can be the enabling technology to make it really work well in a real-world scenario.”

How soon will that be? “It’s already happening,” says Zhao. “It’s not the future. It’s now.”