Reinforcement Learning or Slam Which one is more Promising for Autonomous Driving and Robotics in 2025

I am curious about what do most researchers or practitioners think about "RL or SLAM have more promising future in the application of Autonomous Driving". What are the pros and cons of choosing Reinforcement Learning or Slam?

I have been struggling for a long time when choosing the direction of undergraduate research. Which one is more promising, RL or SLAM? Especially in industrial applications, which direction is easier to find a job? It is difficult to predict the future. Autonomous driving is in the ascendant, and XR is in full swing. A lot of answers advocate that SLAM has good employment prospects while reinforcement learning cannot be implemented. Four years have passed in a blink of an eye, and autonomous driving has begun to have map-free solutions, and the metaverse bubble has burst. Instead, reinforcement learning has become the key technology of large language models. If undergraduates follow the following suggestions, they probably know what it means to look for swords on the boat. Choice is greater than effort, but ironically, how to make the right choice sometimes depends entirely on luck.

It is better to do reinforcement learning for undergraduates. It is not that the difficulty of slam discourages you. Just look at how many people are doing slam on Zhihu alone. Won't you be rolling up the world when you graduate? Most of the people who persuade the questioner to do slam in this post are now trying to solve the slam problem. By the time you have thoroughly studied the basics of SLAM and graduated as a graduate student, the market has already been saturated. Several scenarios that require SLAM on self-driving cars are actually on the verge of mass production. Looking at the odometer performance of several data sets, the error is 1%. Engineers can do it, but why should students do it? The field of reinforcement learning is a blue ocean that has not seen a large number of people entering the market. Isn’t the L2+ strong intelligence route fashionable in recent years? Let’s talk about a small incision in the personalized overtaking experience: Do you change lanes when following a car? When? You can search for how many related results there are on this issue, how many people are working on it, and what research angles there are. Is it possible to adapt to people without machine learning methods? Well, that’s all I have to say.

RL has relatively large deployment barriers not only in the field of autonomous driving, but also in more general real-world problems. Here are some of my personal understandings. From the perspective of RL itself, nowadays, RL is basically refering to Deep Reinforcement Learning. Most of the model-free algorithms studied have almost zero interpretability. It is not a problem to run in games or virtual environments, but there are big problems in applying them in real environments. If you are interested in the practical application of RL, you can pay attention to some model-based methods. Generally, robots deployed in real scenarios have task models. To discuss more academically: most of the problems solved by RL are MDPs and there are many explorations. Some DRL works are said to be able to solve POMDPs, but in fact, they are not much different from the methods of solving MDPs. For example, in the classic DRQN, LSTM is used to compress multiple observations and then output them to the network. Many works have also proved that there is not much difference or even no improvement from DQN. The deeper reason is that for POMDP, only the entire historical set or belief has sufficient statistics. For the historical set (exponentially increasing over time), there is no way to directly feed it into the training network. So now the work is all about compressing multiple observations with RNN and then outputting them, but this is far from the serious belief update, and it is not the same thing in theory. This is also my first point. The interpretability of model-free DRL is basically 0. Everyone is trying various tricks to see if it can run, and post the results when they come out. In real life, there are problems with POMDPs, Dec-POMDPs, and even POSG. These are not easily solved by existing DRL, and the planning algorithms in this regard are also limited to small-scale problems, unlike DRL, which can be expanded on a large scale. There is a lot of work to be done in this regard. I don’t know much about SLAM, so I dare not talk nonsense, but it feels like a relatively mature field.

I found that the students have a little misunderstanding. What I want to express is: both directions are good in work, and those who have energy can try both sides. I don’t mean to discourage either side. SLAM mainly includes lasers, vision, and multi-sensors. The theories behind them are relatively similar, and it is relatively easy and enjoyable to learn. The application in industries are mainly autonomous driving, maps, robots, drones, and AR, and the overall prospects are good. Although there are some controversies in the short-term implementation, it is definitely going to be done in the long run (I find it hard to imagine that autonomous driving or robots will not be done suddenly). It is normal for a company to start a business or go bankrupt, but an industry will not disappear suddenly. Recently, I saw this kind of food delivery robot at a Chinese Restaurant. This one delivers a lot of food and walks more steadily than a person. If the price is right, more and more restaurants will definitely use it in the future, and we will become more and more accustomed to the existence of such robots in our daily lives. Of course, this one currently uses ceiling markers. If SLAM is done well, does it mean that the environment does not need to be arranged, and it is easier to promote? Is the same routine for sweepers and food delivery cars? Will it become cheaper and more popular in the future? These are all obvious. Academic SLAM does have a complete solution based on the existing framework, but is the existing framework the best? Does it meet the needs of real applications? At present, it seems not. Can VSLAM progress from pixels and feature points to the level of "spatial environment understanding"? Can it understand the knowledge of "there is a sofa in my house, there is a chair in front of the sofa, and now I see a chair, so I am in front of the sofa"? "The chair I saw before was yellow, but now I see a black column. I think it should be a leg of the chair, and the color I see has changed because I am standing in the dark." Have such things been realized? Laser SLAM is indeed relatively mature, but there are still many engineering problems. For example, can the food delivery vehicles that have been popular in the past two years build a map of the entire Zhongguancun and deliver food to any two locations? How expensive equipment and manpower are needed to build and maintain such a map? Can the same model be implemented all over the country, or do different roads have their own solutions? Is there a low-cost, large-scale, automated, and fast way to build maps instead of relying on millions of collection vehicles to go back and forth? You see, these problems have to be solved, and it will take several years at least. RL... Although I don't understand RL, and I haven't opened the book I bought (yes, I haven't opened it yet), I roughly know that some complex scenarios are difficult to solve with 10,000 if-else statements, and it's hard to say if I add another 10,000. Some bionic, humanoid robots are also difficult to control their behavior with a bunch of if-else statements (I just heard it, I heard it). RL can use 100 million parameters to exchange for 10,000 if-else, and some scenes can achieve the effect of being indistinguishable from the real thing. Some radical people are definitely willing to accept it. So in summary, I think both sides are actually OK. I am also interested in learning some RL things (I also want to learn cooking and music). There are always more speculators in this world and fewer believers. Although speculation cannot be said to be wrong, in the long run, it is easy to "gather when seeing the signs of victory and flee when seeing the signs of defeat", which is not conducive to the development of the industry. I think when you are young, you might as well think more simply, "believe" in certain things, and move forward boldly. Anyway, there are people who are optimistic and pessimistic about everything. No one can convince anyone. You will know it only after doing it. Things are in the hands of people. Of course, "belief" is not "blind faith", and rationality is still needed.

From the perspective of mass production, SLAM has a much better prospect than reinforcement learning. Reinforcement learning can be used as an experiment, but since planning requires much higher interpretability than perception modules such as image laser, pure reinforcement learning is very weak in this aspect and has poor usability. In addition, reinforcement learning requires negative samples, and autonomous driving cannot obtain these negative samples unless you have a very good (not seen yet) simulation system. Therefore, logical planning and development is the mainstream, and the deep learning strategy that may be used is parameter search. If the questioner is seeking stability, I personally suggest that you choose SLAM.

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.