Current results aren't consistent with our expectations (and previous results):
Walker 2D
![image](https://user-images.githubusercontent.com/5062458/61424639-0ce8ff80-a8c9-11e9-8efe-ab0e3d8a1fef.png)
![image](https://user-images.githubusercontent.com/5062458/61425045-6f8ecb00-a8ca-11e9-9e54-7f56aaf5a230.png)
Walker 3D (fixed variance)
![image](https://user-images.githubusercontent.com/5062458/61187957-4ac9f780-a62d-11e9-9fcc-fc0629c27844.png)
![image](https://user-images.githubusercontent.com/5062458/61425520-0c9e3380-a8cc-11e9-8b1a-64a4ae748e27.png)
Walker 3D (learned variance)
![image](https://user-images.githubusercontent.com/5062458/61187962-574e5000-a62d-11e9-807e-f7cc933b9f4b.png)
- Though the net should have use a symmetric variance, so it's expected that it might not work.
Humanoid
![image](https://user-images.githubusercontent.com/5062458/61426914-fcd51e00-a8d0-11e9-9dfb-3150606c9a9f.png)
Cassie 2D
![image](https://user-images.githubusercontent.com/5062458/61187999-cc218a00-a62d-11e9-8509-01122308a7be.png)