Keyboard humanoid demos

Author: olegklimov

agent_zoo/demo_keyboard_humanoid1.py

@@ -0,0 +1,56 @@
+import gym, roboschool, sys, os
+import numpy as np
+import pyglet, pyglet.window as pw, pyglet.window.key as pwk
+from pyglet import gl
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
+import tensorflow as tf
+
+#
+# This opens a test window (not chase camera), allows to control humanoid using keyboard.
+#
+
+def demo_run():
+    env = gym.make("RoboschoolHumanoidFlagrun-v1")
+
+    config = tf.ConfigProto(
+        inter_op_parallelism_threads=1,
+        intra_op_parallelism_threads=1,
+        device_count = { "GPU": 0 } )
+    sess = tf.InteractiveSession(config=config)
+
+    from RoboschoolHumanoidFlagrunHarder_v1_2017jul import ZooPolicyTensorflow
+    pi = ZooPolicyTensorflow("humanoid1", env.observation_space, env.action_space)
+
+    class TestKeyboardControl:
+        def __init__(self):
+            self.keys = {}
+            self.control = np.zeros(2)
+        def key(self, event_type, key, modifiers):
+            self.keys[key] = +1 if event_type==6 else 0
+            #print ("event_type", event_type, "key", key, "modifiers", modifiers)
+            self.control[0] = self.keys.get(0x1000014, 0) - self.keys.get(0x1000012, 0)
+            self.control[1] = self.keys.get(0x1000013, 0) - self.keys.get(0x1000015, 0)
+
+    obs = env.reset()
+    eu = env.unwrapped
+
+    still_open = env.render("human") # This creates window to set callbacks on
+    ctrl = TestKeyboardControl()
+    eu.scene.cpp_world.set_key_callback(ctrl.key)
+
+    while 1:
+        a = pi.act(obs, env)
+
+        if (ctrl.control != 0).any():
+            eu.walk_target_x = eu.body_xyz[0] + 2.0*ctrl.control[0]
+            eu.walk_target_y = eu.body_xyz[1] + 2.0*ctrl.control[1]
+            eu.flag = eu.scene.cpp_world.debug_sphere(eu.walk_target_x, eu.walk_target_y, 0.2, 0.2, 0xFF8080)
+            eu.flag_timeout = 100500
+
+        obs, r, done, _ = env.step(a)
+        still_open = env.render("human")
+        if still_open==False:
+            return
+
+if __name__=="__main__":
+    demo_run()

agent_zoo/demo_keyboard_humanoid2.py

@@ -0,0 +1,91 @@
+import gym, roboschool, sys, os
+import numpy as np
+import pyglet, pyglet.window as pw, pyglet.window.key as pwk
+from pyglet import gl
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
+import tensorflow as tf
+
+#
+# This opens a third-party window (not test window), shows rendered chase camera, allows to control humanoid
+# using keyboard (in a different way)
+#
+
+class PygletInteractiveWindow(pw.Window):
+    def __init__(self, env):
+        pw.Window.__init__(self, width=600, height=400, vsync=False, resizable=True)
+        self.theta = 0
+        self.still_open = True
+
+        @self.event
+        def on_close():
+            self.still_open = False
+
+        @self.event
+        def on_resize(width, height):
+            self.win_w = width
+            self.win_h = height
+
+        self.keys = {}
+        self.human_pause = False
+        self.human_done = False
+
+    def imshow(self, arr):
+        H, W, C = arr.shape
+        assert C==3
+        image = pyglet.image.ImageData(W, H, 'RGB', arr.tobytes(), pitch=W*-3)
+        self.clear()
+        self.switch_to()
+        self.dispatch_events()
+        texture = image.get_texture()
+        gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MAG_FILTER, gl.GL_NEAREST)
+        texture.width  = W
+        texture.height = H
+        texture.blit(0, 0, width=self.win_w, height=self.win_h)
+        self.flip()
+
+    def on_key_press(self, key, modifiers):
+        self.keys[key] = +1
+        if key==pwk.ESCAPE: self.still_open = False
+
+    def on_key_release(self, key, modifiers):
+        self.keys[key] = 0
+
+    def each_frame(self):
+        self.theta += 0.05 * (self.keys.get(pwk.LEFT, 0) - self.keys.get(pwk.RIGHT, 0))
+
+def demo_run():
+    env = gym.make("RoboschoolHumanoidFlagrun-v1")
+
+    config = tf.ConfigProto(
+        inter_op_parallelism_threads=1,
+        intra_op_parallelism_threads=1,
+        device_count = { "GPU": 0 } )
+    sess = tf.InteractiveSession(config=config)
+
+    from RoboschoolHumanoidFlagrunHarder_v1_2017jul import ZooPolicyTensorflow
+    pi = ZooPolicyTensorflow("humanoid1", env.observation_space, env.action_space)
+
+    control_me = PygletInteractiveWindow(env.unwrapped)
+
+    env.reset()
+    eu = env.unwrapped
+
+    obs = env.reset()
+
+    while 1:
+        a = pi.act(obs, env)
+
+        x, y, z = eu.body_xyz
+        eu.walk_target_x = x + 1.1*np.cos(control_me.theta)   # 1.0 or less will trigger flag reposition by env itself
+        eu.walk_target_y = y + 1.1*np.sin(control_me.theta)
+        eu.flag = eu.scene.cpp_world.debug_sphere(eu.walk_target_x, eu.walk_target_y, 0.2, 0.2, 0xFF8080)
+        eu.flag_timeout = 100500
+
+        obs, r, done, _ = env.step(a)
+        img = env.render("rgb_array")
+        control_me.imshow(img)
+        control_me.each_frame()
+        if control_me.still_open==False: break
+
+if __name__=="__main__":
+    demo_run()

roboschool/cpp-household/roboschool.files

@@ -36,6 +36,8 @@ python-binding.cpp
 ../../agent_zoo/demo_race1.py
 ../../agent_zoo/demo_race2.py
 ../../agent_zoo/demo_pong.py
+../../agent_zoo/demo_keyboard_humanoid1.py
+../../agent_zoo/demo_keyboard_humanoid2.py
 
 glsl/common.h
 glsl/displaytex.frag.glsl