Which one is better to Learn for Autonomous Driving ROS2 or Apollo

Hi, my advice is that you need to all all the fundamental robotics and autonomous Driving related subjects, but Apollo should be prioritized with the core Algorithm. ROS2 is just a framework with scheduling and data communication. The autonomous driving technologies of various companies are nothing more than these mainstream solutions~ Perception-related BEV perception: BEV is the cornerstone of the current mainstream mass production solution. The unified feature expression can directly output multi-task results, such as 3D detection, BEV segmentation, online high-precision maps, etc.;

OCC directly models the occupancy information of the 3D world space, which can effectively handle targets with irregular shapes and unclear semantics (such as road garbage and stones), etc., which is essential for driving/parking;

Monocular focuses on cost-effectiveness and has been widely used in various L2/L3 mass production solutions. Monocular 2D/3D/BEV and OCC also have their own unique development trends;

millimeter wave radar, especially 4D Radar is conquering the L4 high ground. Its distance measurement, speed measurement, angle measurement capabilities and unique perception solutions have allowed Radar to go its own way; Point cloud 3D target detection: Point cloud 3D detection is the basis of LV fusion. It has a very rich technical direction and is more difficult. It is currently mainly used in L4 and true value systems;

The purpose of fusion is to solve problems such as visual occlusion and single-frame instability. There is no doubt that LV fusion is the current mainstream mass production direction; RV fusion: The all-weather operation characteristics of millimeter waves make RV fusion complementary and improve perception performance as a whole;

Lane line detection as As the basic task of autonomous driving, it is gradually developing towards online high-precision maps. This task can be said to be the core of mapless NOA; Transformer: Since ViT swept CV, the SOTA models in autonomous driving have more or less had Transformer, which can be said to be leading the trend; Multi-sensor fusion target tracking: In order to facilitate downstream use, perception needs to obtain the timing information of the target, so as to "string" the obstacles together. Currently, both dynamic and static targets can use tracking to output stably. This technology stack involves single sensor/multi-sensor, Kalman filtering, etc.;

2D target detection is very important in autonomous driving. It still occupies a place in driving, such as traffic light detection, road signs, etc.

Trajectory prediction is an important module connecting perception and regulation. Compared with perception, this direction is smaller in scale, but has higher technical requirements. It is a niche but very popular field;

Regulation and control, as the most downstream module of the entire autonomous driving, directly determines the safety and comfort of autonomous driving. This field has high theoretical requirements and a higher upper limit;

The model cannot be separated from the deployment and optimization of the actual vehicle. This field involves a lot of CUDA programming and TensorRT deployment, which can be said to be engineering The most demanding direction;

simulation as the core of the test link can open up the entire data closed loop, especially combined with the currently popular 3DGS three-dimensional reconstruction, it has great potential;

C++ is the most widely used and most important language in autonomous driving, a solid foundation for mass production and a tool that can never be avoided;

the accuracy of calibration parameters directly affects the application of downstream perception and positioning fusion. This technical direction involves sensors such as Lidar/Radar/Camera/IMU;

camera calibration is the most widely used of all sensors. How to choose the imaging model? How to calibrate the internal and external parameters? How to calibrate binocular & surround view & fisheye? How to choose the calibration board is worth exploring;

End-to-end autonomous driving refers to the direct output of vehicle control signals based on input. Tesla has verified the feasibility of the solution, and China is catching up in strides. It will definitely be the high ground for intelligent driving competition in 2024;

The world model aims to understand autonomous driving from a higher dimension, using a large number of real-life videos to generate future scenes, which can be used to generate simulation data on a large scale, especially Corner Case;

Unlike traditional three-dimensional reconstruction methods, NeRF/3D GS is faster and has better performance. It can also be connected with the simulation link, which is a killer for future data closed loops; Large model and autonomous driving: The combination of large language models and autonomous driving is also a cutting-edge field at present, and autonomous driving may usher in the GPT moment;

For more informtion, you can visit Robotic Search Engine to find latest research papers, AI Agent Marketplace and AI Agent Search Engine to find AI Agents related topics reviews and latest product information.