Fsd12刚吹完一波，平均几公里接管一次
【 在 TimeAndRiver 的大作中提到: 】
:    感觉距离越来越远
:
:  
: ...................
发自「快看水母 于 iPhone 14 Pro」
--
FROM 120.245.23.*

平均12英里接管1次，和结果并不好，这个结论是怎么得来的？和谁对比的？waymo？
我觉得图教授的结论有待商榷
【 在 qtpr 的大作中提到: 】
: 国外独立机构测评fsd v12.5的两个小版本（我印象是12.5.1和12.5.3）的结果并不好，平均每12英里需要人工接管一次。
: 端到端的架构可以明显提升自驾操作的平顺性，即所谓的像老司机，很多人被这种像老司机的感觉欺骗，但它并不能明显提升安全性。
:
--
FROM 123.121.56.*

Tesla's Full Self-Driving (FSD) technology has long been hyped as the next big thing in the automotive world. While Tesla CEO Elon Musk often touts the system's capabilities, independent testing and data have often painted a more sobering picture. Now, a new evaluation by AMCI Testing, an independent automotive testing firm, has further exposed the limitations of Tesla's FSD, raising significant questions about its readiness for widespread deployment.
  
AMCI's testing, conducted over 1,000 miles in a 2024 Tesla Model 3 Performance equipped with Hardware 4 and running FSD versions 12.5.1 and 12.5.3, revealed a disconcerting reality: human intervention was required on average once every 13 miles. This translates to over 75 interventions during the testing period, far exceeding the expectations of even the most critical FSD observers.
  
The findings stand in stark contrast to crowdsourced data, which suggested a much higher average distance between disengagements. While AMCI acknowledges that Tesla's FSD system is impressive in its ability to mimic human-like driving behaviors, especially for a camera-based system, it also warns against the dangers of complacency.
  
Guy Mangiamele, Director of AMCI Testing, cautions that the system's initial "infallibility" can create a false sense of security, leading drivers to take their hands off the wheel or become distracted. This, he emphasizes, is incredibly dangerous, as even professional drivers operating with a testing mindset had to remain vigilant to catch split-second miscalculations by the FSD system.
  
Perhaps even more concerning is the unpredictable nature of FSD's failures. Mangiamele notes that the system can successfully navigate a particular scenario multiple times, only to inexplicably fail the next time around. Whether this is due to a lack of computing power, buffering issues, or shortcomings in the system's assessment of its surroundings remains unclear.
  
AMCI's testing also highlighted persistent failures stemming from basic programming inadequacies. For instance, the system often initiated lane changes towards a freeway exit mere tenths of a mile before the exit itself, hindering its overall functionality and raising doubts about the quality of its underlying programming.
  
As the automotive industry continues its march towards fully autonomous vehicles, Tesla's FSD remains a fascinating yet controversial example of the challenges and complexities involved in achieving this goal. With AMCI planning to release more videos and test future FSD updates, the ongoing scrutiny of this technology is likely to continue, keeping both Tesla and the broader automotive community on their toes.

【 在 leafleter 的大作中提到: 】
: 平均12英里接管1次，和结果并不好，这个结论是怎么得来的？和谁对比的？waymo？
: 我觉得图教授的结论有待商榷
--
FROM 111.167.233.*

和同行比，fsd可能在一定程度上是领先的，但不宜夸大宣传。马斯克本人的某些预期过于乐观。主贴的标题也有夸大之嫌。

自动驾驶关乎安全，应谨慎。

【 在 leafleter 的大作中提到: 】
: 这篇文章本身没什么问题，但是并不能说明FSD不好！
: 如果要更客观的去评价FSD，一个是 跟其他竞争对手的自动驾驶对比，一个是跟他自己的历史进化速度对比，才能看到FSD目前的真实情况。
: FSD虽然离完全自动驾驶还有比较长的距离，但其自动驾驶水平一定是领先的，进化速度也是惊人的。
: ...................
--
FROM 111.167.233.*

降低基本场景下的失败率才是更有意义的，要让用户能对系统的安全边界有信心，而不是找一些极端场景来误导用户产生虚假的信心。

评价一个守门员的水平，不是看他的某几次扑救有多困难和精彩，而是看他是否能尽可能减少低级失误。


【 在 TimeAndRiver 的大作中提到: 】
: 大家如果出不去可以去哔站看那个视频
: 一辆特斯拉被一辆公交车和一辆轿车挡住
: 然后它选择倒车超过去
: ...................
--
修改:qtpr FROM 111.167.233.*
FROM 111.167.233.*