Hi everyone! I would like to control a robotic EE in position only, so I wrote in the actuator part of XML code and all the magic stuff needed. The problem I encounter is that during RL the action is sampled in the ctrlrange but I would like to have the whole joint space while keeping a limited action sample. Is that any way to solve this stuff? Thanks!
THE XML FILE
<compiler angle="radian"/>
<option cone="elliptic">
<flag gravity="disable"/>
</option>
<asset>
<texture name="texplane" type="2d" builtin="checker" rgb1=".2 .3 .4" rgb2=".1 0.15 0.2"
width="512" height="512"/>
<material name="MatGnd" reflectance="0.5" texture="texplane" texrepeat="1 1" texuniform="true"/>
<mesh name="link0_collision" file="stl/panda/collision/link0.stl"/>
<mesh name="link1_collision" file="stl/panda/collision/link1.stl"/>
<mesh name="link2_collision" file="stl/panda/collision/link2.stl"/>
<mesh name="link3_collision" file="stl/panda/collision/link3.stl"/>
<mesh name="link4_collision" file="stl/panda/collision/link4.stl"/>
<mesh name="link5_collision" file="stl/panda/collision/link5.stl"/>
<mesh name="link6_collision" file="stl/panda/collision/link6.stl"/>
<mesh name="link7_collision" file="stl/panda/collision/link7.stl"/>
<mesh name="hand_collision" file="stl/panda/collision/hand.stl"/>
<mesh name="finger_collision" file="stl/panda/collision/finger.stl" scale='1.75 1.0 1.75'/>
<mesh name="link0_visual" file="stl/panda/visual/link0.stl"/>
<mesh name="link1_visual" file="stl/panda/visual/link1.stl"/>
<mesh name="link2_visual" file="stl/panda/visual/link2.stl"/>
<mesh name="link3_visual" file="stl/panda/visual/link3.stl"/>
<mesh name="link4_visual" file="stl/panda/visual/link4.stl"/>
<mesh name="link5_visual" file="stl/panda/visual/link5.stl"/>
<mesh name="link6_visual" file="stl/panda/visual/link6.stl"/>
<mesh name="link7_visual" file="stl/panda/visual/link7.stl"/>
<mesh name="hand_visual" file="stl/panda/visual/hand.stl"/>
<mesh name="finger_visual" file="stl/panda/collision/finger.stl" scale='1.75 1.0 1.75'/>
</asset>
<visual>
<scale framewidth="0.05" framelength="0.8" jointwidth="0.05" jointlength="0.8" actuatorwidth="0.05" actuatorlength="0.8" forcewidth="0.1" contactwidth="0.1"/>
</visual>
<default>
<geom condim="4"/>
<default class="panda">
<joint pos="0 0 0" limited="true" damping="100"/>
<position forcelimited="true" ctrllimited="true" user="1002 40 2001 -0.005 0.005"/>
<default class="visual">
<geom contype="0" conaffinity="0" group="0" type="mesh" rgba=".95 .99 .92 1" mass="0"/>
</default>
<default class="collision">
<geom contype="1" conaffinity="1" group="3" type="mesh" rgba=".5 .6 .7 1"/>
</default>
<default class="panda_finger">
<joint damping="0" armature='5'/>
</default>
</default>
</default>
<worldbody>
<light pos="0 0 1000" castshadow="false"/>
<!-- FLOOR -->
<geom name="ground" pos="0 0 0" size="5 5 10" material="MatGnd" type="plane" contype="1" conaffinity="1"/>
<!-- TABLE -->
<!--geom name="wrk_table" pos="0 0 0.2" type="box" mass="90" size=".2 .2 .2" rgba="0.9 0.9 0.9 1" contype="1" conaffinity="1"/-->
<!-- TARGET -->
<body name="target" pos="0 0 .45">
<!--geom type="cylider" size="0.02 0.05" rgba=".9 0 0 .5" contype="8" conaffinity="8"/-->
<site name="target_site" type="cylinder" size="0.02 0.05" pos="0 0 0" rgba="0.9529411765 0.8 0.03529411765 0.5"/>
</body>
<!-- HAND -->
<body name="panda_hand" pos="0 0 0.8" euler="3.14159265359 0 0">
<joint name="panda_x" pos="0 0 0" type="slide" axis="1 0 0" frictionloss="0" damping="1000"/>
<joint name="panda_y" pos="0 0 0" type="slide" axis="0 1 0" frictionloss="0" damping="1000"/>
<joint name="panda_z" pos="0 0 0" type="slide" axis="0 0 1" frictionloss="0" damping="1000"/>
<joint name="panda_ball" pos="0 0 0" type="ball" frictionloss="0" damping="10"/>
<site name="ee_site" pos="0 0 0" size="0.005, 0.005, 0.005" euler="0 0 -1.57"/>
<inertial pos="0 0 0" euler="0 0 0" mass="1" diaginertia="0.1 0.1 0.1"/>
<geom class="visual" mesh="hand_visual"/>
<geom class="collision" mesh="hand_collision"/>
<body name="panda_left_finger" pos="0 0.02 0.0584" quat="1 0 0 0">
<!--joint name="panda_finger_joint1" axis="0 1 0" type="slide" range="-0.02 0.02" damping="1000" armature="500"/-->
<geom class="visual" mesh="finger_visual"/>
<geom class="collision" mesh="finger_collision" mass="1"/>
</body>
<body name="panda_right_finger" pos="0 -0.02 0.0584" quat="1 0 0 0">
<!--joint name="panda_finger_joint2" axis="0 -1 0" type="slide" range="-0.02 0.02" damping="1000" armature="500"/-->
<geom quat="0 0 0 1" class="visual" mesh="finger_visual"/>
<geom quat="0 0 0 1" class="collision" mesh="finger_collision" mass="1"/>
</body>
<!-- COMPONENT -->
<body name="component" pos="0 0 0.18">
<geom type="cylinder" size="0.02 0.05" mass="0.1" margin="0.001"/>
<site name="component_site" pos="0 0 0.05" rgba="0.9529411765 0.8 0.03529411765 1"/>
</body>
</body>
</worldbody>
<actuator>
<position name="panda_x" joint="panda_x" class="panda_finger" gear="1" kp="100000" forcerange="-50 50" ctrllimited="true" ctrlrange="-0.01 0.01"/>
<position name="panda_y" joint="panda_y" class="panda_finger" gear="1" kp="100000" forcerange="-50 50" ctrllimited="true" ctrlrange="-0.01 0.01"/>
<position name="panda_z" joint="panda_z" class="panda_finger" gear="1" kp="100000" forcerange="-50 50" ctrllimited="true" ctrlrange="-0.01 0.01"/>
<position name="panda_ball_1" joint="panda_ball" class="panda_finger" kp="10000" gear="1 0 0 0 0 0" forcerange="-50 50" ctrllimited="true" ctrlrange="-0.03 0.03"/>
<position name="panda_ball_2" joint="panda_ball" class="panda_finger" kp="10000" gear="0 1 0 0 0 0" forcerange="-50 50" ctrllimited="true" ctrlrange="-0.03 0.03"/>
<position name="panda_ball_3" joint="panda_ball" class="panda_finger" kp="10000" gear="0 0 1 0 0 0" forcerange="-50 50" ctrllimited="true" ctrlrange="-0.03 0.03"/>
</actuator>
<sensor>
<force name="ee_force_sensor" site="ee_site"/>
<torque name="ee_torque_sensor" site="ee_site"/>
</sensor>
THE ENVIRONMENT
import numpy as np
from gym import utils
from gym.envs.mujoco import mujoco_env
class PandaEnv(mujoco_env.MujocoEnv, utils.EzPickle):
def __init__(self):
utils.EzPickle.__init__(self)
mujoco_env.MujocoEnv.__init__(self, "/home/bara/doc/rlkit/generic/panda.xml", 2)
def step(self, a):
# STEP
self.do_simulation(a, self.frame_skip)
#DISTANCE
xpos_component = self.get_body_com("component")
xpos_target = self.get_body_com("target")
dist = xpos_component - xpos_target
#REWARD
reward_dist = -np.linalg.norm(dist)
reward = reward_dist # More contributions to rewards may be added
# print("REWARD: " + str(reward))
# print("ACTION: " + str(a))
ob = self._get_obs()
done = False
return ob, reward, done, dict(reward_dist=reward_dist)
def viewer_setup(self):
self.viewer.cam.trackbodyid = 0
def reset_model(self):
qpos = np.zeros(self.model.nq)
qvel = np.zeros(self.model.nv)
self.set_state(qpos, qvel)
return self._get_obs()
def _get_obs(self):
return np.concatenate(
[
self.get_body_com("component"),
self.get_body_com("target")
]
)
Hi everyone! I would like to control a robotic EE in position only, so I wrote in the actuator part of XML code and all the magic stuff needed. The problem I encounter is that during RL the action is sampled in the ctrlrange but I would like to have the whole joint space while keeping a limited action sample. Is that any way to solve this stuff? Thanks!
THE XML FILE
THE ENVIRONMENT