- Discusses a fully automatic method for generating gaits and morphologies for legged locomotion
- Given a shape, it will determine an effective gait for locomotion
- It can also adapt a bodies morphology to be optimized for a particular locomotion task
- They do not require shaping (an initial gait proposal)
- Method combines derivative-based and derivative free search, with the first probably suited for fine-grained optimization and the latter for dealing with discontinuities
- Show the approach with bodies that have 2, 4, and 5 legs
- There are a number of references here worth checking out (mostly in section 2)
- Simulation is more physically accurate than my stick figures as it simulates elasticity that results from actual muscles, also more accturately models skeleton and muscle activiation
- Models active and passive forces
- Reward function attempts to encourage realistic gaits by penalizing high joint velocities, large torques, and large amounts of head movement, for example
- Optimization is based on a variant of Covariance Matrix Adaptation (CMA)
- The version they use, called basin-CMA, is designed to take advantage of derivatives when available as well as operate under constraints required by the skeleton and other criteria
- It is called basin-CMA because it operates in the basins of attraction; projects the original space into the basin
- There are some other tricks, such as operating in the projected space as opposed to the full space
- The also say basin-CMA outperformed top methods on the problems from the 2006 Congress on Evolutionary Computation test set, while also being computationally cheaper
- Separates optimization of foot placement and then does the rest of the body to match that to make the process simpler
- Experiments are run on skeletons with 10 to 32 DOFs, fully in 3D, generally takes 50 iterations of optimizaiton
- Very nice paper