华为乾崑ADS智驾,传统自动驾驶技术最后的余晖
我一直关注自动驾驶的技术路线之争。其中,华为的声音可以说是最大的,从ADS 1.0一路迭代到最新的乾崑ADS 4.0,声势浩大 。最近他们推出了所谓的WEWA架构,即世界引擎World Engine加上世界行为模型World Action Model 。这套系统被华为包装成先进的,甚至是端到端的架构 。

但作为一名自封的《原教旨端到端主义者》,我必须指出,华为的这套东西,本质上是上一代的技术。
华为确实说过自己是端到端,但这个端到端的程度值得商榷。如果说小米的技术路线实现了80%的端到端,那么华为在我看来,最多只有60%。原因很简单,就出在它的World Engine上。

WEWA架构的工作方式,是先通过车辆采集海量真实路况数据,然后将这些数据上传到云端,在World Engine这个所谓的数字孪生世界里进行映射、重建和增强 。云端的AI在这个虚拟世界里学习各种极端场景,训练出一个模型,再下发到车端的World Action Model去执行 。

这个流程存在一个明显的分离步骤。数据并非直接从传感器输入到统一的大模型,然后直接输出控制指令。中间横亘着一个庞大的世界引擎。它是一个数据处理和仿真中心,是一个中间环节。这种先构建一个虚拟世界,再让模型在虚拟世界里学习的做法,在我看来,这是传统模块化自动驾驶思路的终极形态,而不是真正的端到端革命。我不认可小米的端到端,更不认可华为的。
当然,我必须承认,华为这套上一代的技术,在当下的实际效果并不差。它通过海量的工程投入和精细的规则优化,在高速等结构化道路上表现得十分稳定,甚至可以说是顶级水准。这套系统的体验打磨得非常成熟,这是事实。
但问题在于,非纯粹端到端的系统,其技术上限是清晰可见的。它的天花板很低。当系统发展到一定程度后,想再提高一点点的驾驶水平,需要付出的成本和努力将是指数级增长的。每一个细微场景的优化,都需要对感知、规划、决策等多个模块进行繁琐的调整和验证,牵一发而动全身。
这正是我为华为智驾未来感到担忧的地方。
看看竞争对手,小鹏和特斯拉正在纯粹端到端的道路上快速进步。他们致力于用一个统一的模型解决问题,虽然早期可能磕磕绊绊,但其上限极高,迭代速度会越来越快。而华为,则需要持续维护和优化WEWA这一套尾大不掉的系统。这套系统的维护成本极高,因为它需要庞大的云端算力和数据团队来支撑世界引擎的运转 。然而,如此巨大的投入,换来的可能只是性能上非常有限的进步。

华为的乾崑智驾,更像是传统自动驾驶技术路线最后的余晖。它用极致的工程能力将旧思路推向了顶峰,但终究只是落日的辉煌。真正的未来,属于更简洁、更统一、更彻底的端到端架构。继续在旧的道路上投入,只会离未来越来越远。

值友8827439162
Unlike peer L2 systems tested by ODI, Autopilot presented resistance when drivers attempted to
provide manual steering inputs. Attempts by the human driver to adjust steering manually resulted
in Autosteer deactivating. This design can discourage drivers’ involvement in the driving task. Other
systems tested during the PE and EA investigation accommodated drivers’ steering by suspending
lane centering assistance and then reactivating it without additional action by the driver.
Notably, the term “Autopilot” does not imply an L2 assistance feature, but rather elicits the idea of drivers not being in control.
校验提示文案
花开花落花自知
校验提示文案
值友8827439162
下限是要命的安全问题。。
校验提示文案
电钮没电
校验提示文案
值友8827439162
校验提示文案
值友8827439162
involving hazards that would have been visible to an attentive driver. Before August 2023, ODI reviewed 956 total crashes where Autopilot was initially alleged to have been in use at the time of, or leading up to those crashes.
校验提示文案
电钮没电
校验提示文案
值友8827439162
校验提示文案
花开花落花自知
校验提示文案
值友8827439162
involving hazards that would have been visible to an attentive driver. Before August 2023, ODI reviewed 956 total crashes where Autopilot was initially alleged to have been in use at the time of, or leading up to those crashes.
校验提示文案
值友8827439162
Unlike peer L2 systems tested by ODI, Autopilot presented resistance when drivers attempted to
provide manual steering inputs. Attempts by the human driver to adjust steering manually resulted
in Autosteer deactivating. This design can discourage drivers’ involvement in the driving task. Other
systems tested during the PE and EA investigation accommodated drivers’ steering by suspending
lane centering assistance and then reactivating it without additional action by the driver.
Notably, the term “Autopilot” does not imply an L2 assistance feature, but rather elicits the idea of drivers not being in control.
校验提示文案
值友8827439162
下限是要命的安全问题。。
校验提示文案