From de959289afd8c8a3220b28c9d497ecde3c15d6ea Mon Sep 17 00:00:00 2001
From: Joe Booth <joe@joebooth.com>
Date: Wed, 15 Aug 2018 00:22:32 -0700
Subject: [PATCH] document experiments

---
 README.md | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/README.md b/README.md
index 88e3374..0eceb88 100644
--- a/README.md
+++ b/README.md
@@ -1,10 +1,31 @@
 # ActiveRagdollControllers
 Research into controllers for 2d and 3d Active Ragdolls (using MujocoUnity+ml_agents)
 
+----
+
+#### Contributors
+* Joe Booth ([SohoJoe](https://github.com/Sohojoe))
+
+----
+#### Download builds (Mac, Windows): [see Releases](https://github.com/Sohojoe/ActiveRagdollControllers/releases)
+--- 
+### Controller004
+![Controller004](images/Controller004.13-10m.gif)
+* **Type:** Discrete 2D
+* **Build (MacOS)** TODO [v0.004](https://github.com/Sohojoe/ActiveRagdollControllers/releases/tag/v0.003) **Playable**
+* **Actions:** No-op, Forward, Backwards, Jump, Jump+Forward, Jump+Backwards
+* **Controls:** Left arrow, Right arrow, Spacebar
+* **Mujoco Model:** DeepMindHopper
+* **Hypostheis**: Use discreate + random trainer to create human controller.
+* **Outcome:** 
+  * **SUCCESS** - contriol feels responsive
+  * ... Has emerging functionality - i.e. tap left for small step, swap direction in air 
 
 
 ### Controller003
+![Controller003](images/Controller003.gif)
 * **Type:** Continuous 2D
+* **Build (MacOS)** [v0.003](https://github.com/Sohojoe/ActiveRagdollControllers/releases/tag/v0.003)
 * **Actions:** Forward / Backwards
 * **Mujoco Model:** DeepMindHopper
 * **Hypostheis**: Use an adversarial hierarchical trained agent as the controller which gets the inverse reward of the locomation agent on a slower time step. The idea is that it will push the locomoation agent to focus on its weakest areas. 
@@ -15,7 +36,9 @@ Research into controllers for 2d and 3d Active Ragdolls (using MujocoUnity+ml_ag
 
 
 ### Controller002
+![Controller002](images/Controller002.gif)
 * **Type:** Continuous 2D
+* **Build (MacOS, Windows)** [v0.002](https://github.com/Sohojoe/ActiveRagdollControllers/releases/tag/v0.002)
 * **Actions:** Forward / Backwards
 * **Input:** Unity Axis input (left/right or a/d or joystick)
 * **Mujoco Model:** DeepMindHopper
@@ -35,7 +58,9 @@ Research into controllers for 2d and 3d Active Ragdolls (using MujocoUnity+ml_ag
   * Controller002InputAgent.cs - 
 
 ### Controller001
+![Controller001](images/Controller001.gif)
 * **Type:** Discrete 2D
+* **Build (MacOS)** [v0.001](https://github.com/Sohojoe/ActiveRagdollControllers/releases/tag/v0.001)
 * **Actions:** Forward / Backwards
 * **Mujoco Model:** DeepMindHopper
 * **Hypostheis**: It should be simple to train a backwards / forwards by giving the agent a +1 / 1 velocity target which feeds the reward function.