I am a master student in Germany and one of our professors is doing robotics research but with a different paradigm - the behavior-based robotics based on Subsumption Architecture developed by prof. Rodney Brooks from MIT in 1986. I was looking online for more info about this topic but could not find much, compared to the popular nowadays Deep learning techniques. Do you think behavior-based robotics is a fruitful research direction? Does it have any future? At this moment I am inclined to think that Deep Learning and Reinforcement Learning are the ways to add intelligence to robotic systems.

  • $\begingroup$ Current AI trends are just that..**trends**...because lots of people are working on them, it simply advances quicker, and has very very good results...there are lots of reasons that DL and RL are good, but lots that are not good. B.b. R is a field that not so many are working on, that doesn’t mean it has no future....it’s just a slower one that hasn’t had as many break throughs...maybe your prof will have one or two? $\endgroup$ Commented Aug 28, 2019 at 15:43

2 Answers 2


I have not seen that many people working on behavior-based robotics. I am not sure even if Rodney Brooks still works on the subject.

Deep Learning: the approach is good for problems without a model. It is not to-go approach for many robotics problems. Because data collection, labeling, training, testing can be very costly. Even if you are interested in paying the price, sometimes there may not be enough data. If you want to land a rocket to a moving platform, probably you wouldn't want to use deep learning. :)

Reinforcement Learning: the approach does not work in real-life applications. It is costly since you need to do a lot experiments. Most of the times it is not easy to interpret what your agent actually learned.

=== purely subjective statements begin ===

My friendly advice: stay away from esoteric research subjects. I've seen some university professors who have a thing for obscure topics. These professors usually are the ones who passed the critical phase of their careers (pre-tenure) and like to experiment different topics at the expense of their students' career. It is highly likely that the outcome of the experiment will not make the professors more or less famous. They win some small grants and lure some young students to work on esoteric topics as if the subject were very important. You should ask yourself what you are going to achieve after you successfully study "xyz" topic. Is it going to help you find a job in industry or is it going to get you your own research agenda? Do not study a subject just because your professor has funding and he/she finds that topic interesting.

=== purely subjective statements end ===

  • $\begingroup$ What paradigms do you think is best to study for specific robotic AI? $\endgroup$ Commented Aug 28, 2019 at 19:08
  • 1
    $\begingroup$ I am not an authority in robotics to decide which paradigm is best. My expertise is in decision making (route planning, mission planning, behavior planning, motion and path planning), estimation and control. I'd say if you cannot interpret the output of an algorithm (such as in deep learning, reinforcement learning), that algorithm will never fly on a real robotic platform. Simple A*, PID, Kalman filters will be better approach than most of the complex ML algorithms. $\endgroup$
    – Octavius
    Commented Aug 30, 2019 at 5:40

Deep Learning and behavior based robotics as well were developed as the opposite to something which was common before. In the 1980s, Brooks developed the subsumption architecture as an alternative to the former AI planning technique which was widespread discussed in the 1970s together with LISP symbolic systems. Behavior based robotics says in short, that there is no need for an abstract representation of the world which is formalized in the STRIPS language, instead the robot has to react to incoming sensor perception.

Deep Learning were also developed as a valid alternative to something known before. The main idea behind machine learning is, that there is no need to create sourcecode like in a programming project. It's enough to train neural networks with example data and the programming part of a robotics project is low or not there. A certain paradigm becomes successful, if a larger audience has the same understanding of how technology should be realized. The reason why deep learning is more attractive to a mainstream audience than behavior based robotics, is because the number of people who are trying to escape from symbolic AI is low, but the amount of people who doesn't understand why they should program is large.

Deep learning can be summarized to the idea, not to write a single line of code, but letting the combination of a GPU plus backpropagation algorithm do all the work. This sounds attractive to newbies in the domain of AI because writing millions lines of code is a complicated issue.

  • $\begingroup$ So the classical way to program is still far more suited for robotics AI than DL/RL? $\endgroup$ Commented Aug 28, 2019 at 19:12
  • $\begingroup$ It's hard to judge about neural networks with a neutral point of view, because the development is in an early stage. Most of the papers were published in the last 5 years and the idea of utilizing machine learning for solving AI problems makes sense. What we can say for sure is, that deep learning was equal to a revolution. $\endgroup$ Commented Aug 28, 2019 at 19:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.