WingNews logo WingNews GitHub [2]
top | item 11864527

Ask HN: What are the modern approaches to robotic control?

43 points| bistable | 9 years ago | reply

I'm eager to find references on modern approaches to robotic control.

The coordinated robots from Boston Dynamics (acquired by Google) impress the most, with their ability to react quickly to perturbation and move competently over difficult terrain.

I assume they are not solving inverse kinematics problems continuously, and also assume that much of their progress was made before the deep learning revolution, so maybe they don't even use neural networks either.

So are they using any reinforcement learning? I have no idea so would appreciate any quick summaries or pointers to relevant information.

What other companies are working in robotics at the same level as Boston Dynamics?

Approaches to robot perception, path planning (e.g. A*), and environmental mapping (e.g. SLAM) are also very interesting, but not the topics I'm interested in for this question.

20 comments

order
[+] ModernMech|9 years ago|reply
You say you're not interested in path planning, but path planning and control are really on the same spectrum, which is motion planning. For instance, where to next put big dog's leg is a control problem, but also a path planning problem (how do I get the leg there? Where is the best place to put the leg? Does the leg collide with anything?). If you can't answer these questions and are blindly putting your leg wherever, you're not in control of anything.

Here's a good paper on RL in robotics: http://www.ias.tu-darmstadt.de/uploads/Publications/Kober_IJ...

RL is not widely used for control, but it has yielded some impressive results. In my experience I had a highly dynamic system for which I built a hand-tuned model for motion planning. I also built a RL model and trained it using the hand-tuned model. The RL model performed more than 50% better than my very best efforts.

Also, some of my favorite textbooks:

Principles of Robot Motion by Choset et. al.

Statistical Robotics by Fox, Burgard, and Thrun

Linear Systems Theory by Hespanha

[+] antiuniverse|9 years ago|reply
Not my field, but in case this link is helpful:

"Planning Algorithms" by Steven M. LaValle

http://planning.cs.uiuc.edu/

>This book presents a unified treatment of many different kinds of planning algorithms. The subject lies at the crossroads between robotics, control theory, artificial intelligence, algorithms, and computer graphics. The particular subjects covered include motion planning, discrete planning, planning under uncertainty, sensor-based planning, visibility, decision-theoretic planning, game theory, information spaces, reinforcement learning, nonlinear systems, trajectory planning, nonholonomic planning, and kinodynamic planning.

[+] falcolas|9 years ago|reply
For more generalized information on how you bring a machine from state A to state B, you may want to read up on Control Systems [0], one of which is the PID controller [1]. It's used in everything from industrial water heaters to quadcopters. It wouldn't surprise me to see them also being used at some level with more advanced robotics.

[0] https://en.wikipedia.org/wiki/Control_system [1] https://en.wikipedia.org/wiki/PID_controller

[+] stevendhansen|9 years ago|reply
PID is a generalized control algorithm that performs poorly unless your system is entirely linear. It is overwhelmingly used in industry, however, because it is simple, and anyone can drop in a PID block and adjust three constants to get semi-OK response without understanding anything about the dynamics of the system they are controlling.

For serious control systems you can look up state feedback control, model predictive control, and nonlinear systems control (to name a few very broad categories out of many possible options).

[+] srlake|9 years ago|reply
It's been a few years since studying in this area, but one approach we used to use for modelling and control of manipulators (mostly industrial robotic arms) was Screw Theory.

Here's a recent paper using one of these approaches: http://www.intechopen.com/books/international_journal_of_adv...

One approach that could be used for a system like Big Dog would be a feed-forward control loop with kinematic/dynamic modelling of the robot. These approaches use knowledge of the system dynamics to predict the output based on changing inputs or disturbances.

[+] glial|9 years ago|reply
You'll really be missing out if you don't look into the work of Emo Todorov:

http://homes.cs.washington.edu/~todorov/

He's done some amazing work recently in helping solve problems surrounding movement planning (not really the same as path planning like A*), so you can tell a robot to do something general like "stand here", and it treats the movement planning as an optimization problem and the dynamics of its 'body' and physics as constraints. The resulting behavior is eerily lifelike.

[+] sharemywin|9 years ago|reply
they use something called "force control" which I think uses a spring to act as some kind of tendon.

http://news.nationalgeographic.com/news/2003/03/0311_030313_...

http://web.mit.edu/spotlight/archives/troody.html

Mentions some of the engineers at Boston Dymanics including the inventor of troody. https://prezi.com/tn5kji7ehk3u/prezi-project/

[+] sharemywin|9 years ago|reply
more info here:

http://news.mit.edu/2001/dinosaur

We use series elastic actuators, or springs connected to motors, and low-stiffness control, so the joints are looser, and more biological."

HOW IT WORKS Troody, which weighs about ten pounds, has 16 joints and 36 sensors. "Every joint has a position and force sensor," Mr. Dilworth said. The robot also has a vestibular system -- the equivalent of an inner ear -- that it uses for balancing, and an onboard computer that automatically runs a walking control algorithm.

[+] npmanor|9 years ago|reply
If you're interested in walking, the Florida Institute for Human and Machine Cognition (ihmc: http://robots.ihmc.us/) came in second place at the DARPA Robotics Challenge using force control and capture point dynamics. They do primarily software, but their walking is quite stable.