I have a grid of rectangles acting as blocks. The robot traverses through the inter-spaces between these consecutive blocks. Now I have sensor data streaming in representing Right and left wheel speeds. Based on the differences in the speeds of the left and right wheels, I infer the robot's position and the path it has threaded. I get the associated individual segments of the total distance when it travels straight, left, or right.

These distances are a function of the actual speed of the robot and the time interval elapsed before the end of that activity. These computed distances for the segments though don't map and fit-in well when projected on the grid layout of the environment. The segments are rather not adhering to the boundary limitations.

I wanted to know if I can use RL to force the calculated distances to fit in with the layout given certain knowledge (or conditions, if you will): the start and end position of the robot and the inter-space distances.

If not RL, do you know how can I solve this problem? I suspect my function computing the distances is off and wondering if RL can help me figure out the right mapping of sensor data to the path traveled adhering to the grid layout dimensions.

enter image description here

If you consider the illustration above you will notice S, D, and D' signifying the starting position, the true destination, and the destination location computed by adding together the calculated distances for each of the segments representing right(r), left(l) and straight(s) along the path towards the destination. Inter-space length is given 7m and dimensions of the blocks are (27m x 15m). If you look at the data presented on the left side you will notice 18m left and consecutive 24m right represents the activity, in the grid, as the passage through the blocks. Granted -- perhaps the car negotiates the edges and corners through this passage in a protracted left(l) and right(r) movements, without necessarily going straight(s) straddling and linking the turns as one would expect.

The question arises, however, when taken into account these individual segment lengths and stitch them together you end up in a destination, not in the ballpark range of the expected value. How can we design this problem so as to employ RL methods to, sort of, impose these grid dimensional constraints on this distance calculation methodology to yield better results? Or, probably best to re-imagine the whole problem so it is amenable to the application of RL.

Any advice/ insights would be appreciated.

  • $\begingroup$ When I compute the segment distances as a function of speeds of the wheels, I get the values not really conforming to the dimensions of the grid blocks. I suspect the function employed is a little off (plausibly compounded by the rate of data streaming) and want to explore RL for its ability to learn from the given conditions such as the start, the destination and the grid inter-spaces in getting these distances in segments, the cars inferred to be undertaking, sit nice with the grid organization. $\endgroup$ – user007 Oct 9 '18 at 17:25
  • $\begingroup$ Thanks, I think I already understood that from your question. the function is not correct, and you want to have a method to correct it. However, you are not explaining how the discrepancy is being detected. How do you know the value is wrong, and is this something that could be automated? How are you detecting it - by watching the robot? How do you intend to measure the error in the function for learning to work? A learning process needs data $\endgroup$ – Neil Slater Oct 9 '18 at 17:42
  • $\begingroup$ Thank you @NeilSlater. I tried illustrating the problem by adding the picture and the explanation in words. I hope it addresses some of the questions you raised and helps you understand the problem better. $\endgroup$ – user007 Oct 9 '18 at 18:41
  • $\begingroup$ I intend to get data on the start point and the true destination for a few cycles the car traversals. I shall then look for the dynamics and the patterns in the movement of the destination so as to simulate data, by extrapolating, for some more cycles as time progresses. So, we can suppose we have the true destination for quite a few cycles to enable RL to learn to correct the erroneous function that is throwing the output significantly off. I was wondering if this problem can be solved using RL for a desired effect. $\endgroup$ – user007 Oct 9 '18 at 18:51
  • $\begingroup$ Thanks for adding the picture. It makes much more sense to me now. $\endgroup$ – Neil Slater Oct 9 '18 at 18:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.