There are, as always, a large number of different directions of future work. They will be divided up into the following two categories.Improvements on Current Work:
- The model needs to be vetted. This means comparing its results to the complete model and the output of SD/FAST, as well as comparing the chosen torque and dissipative constant values to those on actual motors.
- The searches could use some fleshing out. There is likely a middle of the road value for lambda2 that will give better results. Many of the current controllers move quite noticeably out of position in the current scheme.
- Also, it may be worth reintroducing the pulse count term in the cost. It was of no value in the early stages of annealing, but may prove more valuable in the final steps.
Further Exploration
- Given a way to do calculations on the full (i.e. non-simplified) model in reasonable time, it would be very interesting to optimize over the number and positions of the motors (since with real links, each pivot need not have a motor). Pulse shapes for the motors other than square waves (in either model) could also prove very interesting.
- The spring cable model would likely work quite well, though simulation would be slow without an integrator written in C.
- While the current model was trained to do simple tasks, it does not know how to generalize. A more interesting task than gripping an object in a single position would be to teach the tail to grip an object in an arbitrary position, given the object radius and some simple feedback (perhaps distance from pivots to object).