Tesla FSD 自动驾驶软件升级版本说明书解析三（Beta v10.13 Release Notes）

Tesla FSD软件的版本说明书写得十分技术流，里面会详细列举新增了什么feature、修复了什么bug、采用了什么方法（算法还是数据等）提升了多少指标之类的信息，有助于从业人员了解Tesla的底层算法技术和功能设计逻辑，值得阅读研究。

Telsa FSD于2022年7月18日正式发布了Beta v10.13版本（目前仅面向Tesla员工发布。Available only in Canada, USA），适用于Model 3, Model Y, Model S and Model X。此次属于大版本更新，所以Release notes一共列写了16条，让我们逐条解析看看：

1. Improved decision making for unprotected left turns using better estimation of ego's interaction with other objects through the maneuver.

利用自车机动获得更好的交互估计（与他车等），提升了无保护左转的决策效果

2. Improved stopping pose while yielding for crossing objects at "Chuck Cook style" unprotected left turns by utilizing the median safety regions.

对于Chuck Cook大叔（油管上比较有名的Tesla Beta测试博主）测试的那种极具挑战的无保护长左转（中间需停泊等待，地点在美国Jacksonville, Florida），通过利用中间安全区域提升了让道时的临时停车姿态

3. Made speed profile more comfortable when creeping for visibility, to allow for smoother stops when protecting for potentially occluded objects.

优化怠速缓行时的速度控制策略，当需要避让突然出现的遮挡目标时可以使得刹车更加平滑

4. Enabled creeping for visibility at any intersection where objects might cross ego's path, regardless of presence of traffic controls.

在任意十字路口（无论是否存在交通管制），都能够开启怠速缓行以观察路口情况，应对前方可能出现的横穿目标

5. Improved lane position error by 5% and lane recall by 12% with a full vector lanes neural network update. Information bottlenecks in the network architecture were abated by increasing the size of the per-camera feature extractors, video modules, and internals of the GPT attention modules.

降低了车道位置误差5%，提高车道线召回12%，通过升级全矢量化的车道线神经网络来实现。该网络增大了每一路视频输入的特征提取器、时序模块和注意力模块内部构件的分辨率以减少信息瓶颈的影响

6. Improved lane position error of crossing and merging lanes by 22% by adding long-range skip connections and a more powerful trunk to the network architecture.

将交叉及合并类型的车道线位置误差降低了22%，通过在网络结构中增加长距离的残差连接和更强大的主干网络

7. Improved pedestrian and bicyclist velocity error by 17%, especially when ego is making a turn, by improving the onboard trajectory estimation used as input to the neural network.

降低行人和骑行人的速度估计误差17%，特别是在自车转弯时，通过优化用作速度估计网络输入的自车轨迹预测效果实现

8. Improved animal detection recall by 34% and decreased false positives by 8% by doubling the size of the auto-labeled training set.

提升动物检测召回34%、减少误报8%，通过将自动标注的训练数据量翻一番

9. Improved detection recall of far away crossing vehicles by 4% by tuning the loss function used during training and improving label quality.

提升远距离横穿车辆的检测召回4%，通过优化训练损失函数及提升训练数据的标注质量

10. Improved the "is parked" attribute for vehicles by 5% by adding 20% more examples to the training set.

提升车辆“是否停泊”状态识别精度5%，通过增加20%的训练集

11. Upgraded the occupancy network to detect dynamic objects and improved performance by adding a video module, tuning the loss function, and adding 37k new clips to the training set.

升级了视觉障碍物占据栅格预测网络可以检测动态障碍物，同时增加视频时序模块、优化训练损失函数和增加37k训练数据片段提升了网络效果

12. Reduced false slowdowns around crosswalks by better classification of pedestrians and bicyclists as not intending to interact with ego.

降低了人行横道附近的误刹，通过优化行人/骑行人的属性分类使得这两类目标与自车的碰撞预测更精准

13. Reduced false lane changes for cones or blockages by preferring gentle offsetting in-lane where appropriate.

适时选择轻微偏离车道中心线行驶，从而减少因躲避锥形桶、不规则障碍物等做出的大幅度变道带来的不适感

14. Improved in-lane positioning on wide residential roads.

提升了在开阔居民区道路上的车道内定位精度

15. Improved object future path prediction in scenarios with a high yaw rate.

在自车偏航角较大的场景提升了对目标的轨迹预测精度

16. Improved speed limit sign accuracy on digital speed limits by 29%, on signs with difficult relevance by 23%, on 3-digit speeds by 39%, and on speed limit end signs by 62%. Neural network was trained with 84% more examples in the training set and with architectural changes which allocated more compute in the network head.

提升了限速标志牌的感知精度：对数字限速标志的精度提升了29%、对难以关联的标志精度提升23%、对3位数的速度识别精度提升39%、对限速结束标志的识别精度提升62%。该网络新增了84%的训练数据、网络架构上任务头部分分配了更多的计算量

       综上所述我们可以看到此次升级涉及到的模块归纳如下：

感知（perception）：5（车道线）、6（车道线）、7（速度估计）、8（动物检测）、9（远处横穿车辆）、10（是否停泊）、11（视觉障碍物占据栅格）、12（行人/骑行人分类）、16（限速标志牌）

预测（prediction）：15

定位（positioning）：14

规控（PNC）：1、2、3、4、13