File size: 11,926 Bytes
e479c46 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 | """
Use Simulated annealing to optimize the stiffness and damping of the robot arm and minimize 6d pose errors when open-loop unrolling demonstration trajectories.
"""
import argparse
import multiprocessing as mp
import os
import pickle
import numpy as np
from simulated_annealing import sa
from transforms3d.euler import euler2axangle
from transforms3d.quaternions import axangle2quat, quat2axangle, quat2mat
def calc_pose_err_single_ep(episode, arm_stiffness, arm_damping, robot, control_mode):
import gymnasium as gym
import mani_skill2_real2sim.envs
from sapien.core import Pose
assert robot in ["google_robot_static", "widowx"]
if robot == "google_robot_static":
# append dummy stiffness & damping for the camera links in google robot, which do not affect the results
arm_stiffness = np.concatenate([arm_stiffness, [2000, 2000]])
arm_damping = np.concatenate([arm_damping, [600, 600]])
if robot == "google_robot_static":
sim_freq, control_freq = 252, 3
elif robot == "widowx":
sim_freq, control_freq = 500, 5
env = gym.make(
"GraspSingleDummy-v0",
control_mode=control_mode,
obs_mode="rgbd",
robot=robot,
sim_freq=sim_freq,
control_freq=control_freq,
max_episode_steps=50,
)
# set arm stiffness and damping
env.agent.controller.controllers["arm"].config.stiffness = arm_stiffness
env.agent.controller.controllers["arm"].config.damping = arm_damping
env.agent.controller.controllers["arm"].set_drive_property()
tcp_poses_at_base = []
gt_tcp_poses_at_base = []
_ = env.reset()
env.agent.robot.set_pose(Pose([0, 0, 1]))
def get_tcp_pose_at_robot_base():
tcp_pose_at_robot_base = env.agent.robot.pose.inv() * env.tcp.pose
return tcp_pose_at_robot_base
# unroll the episode and record the tcp (end-effector tool-center-point) poses
for step_id, episode_step in enumerate(episode):
gt_tcp_xyz_at_base = episode_step["base_pose_tool_reached"][:3]
gt_tcp_wxyz_at_base = episode_step["base_pose_tool_reached"][3:]
gt_tcp_pose_at_robot_base = Pose(p=np.array(gt_tcp_xyz_at_base), q=np.array(gt_tcp_wxyz_at_base))
tcp_pose_at_robot_base = get_tcp_pose_at_robot_base()
if step_id == 0:
# At the beginning of episode, set the end-effector pose to be the same as the first observation
controller = env.agent.controller.controllers["arm"]
cur_qpos = env.agent.robot.get_qpos()
init_arm_qpos = controller.compute_ik(gt_tcp_pose_at_robot_base)
cur_qpos[controller.joint_indices] = init_arm_qpos
env.agent.reset(cur_qpos)
tcp_pose_at_robot_base = get_tcp_pose_at_robot_base()
tcp_poses_at_base.append(tcp_pose_at_robot_base)
gt_tcp_poses_at_base.append(gt_tcp_pose_at_robot_base)
gt_action_world_vector = episode_step["action_world_vector"]
gt_action_rotation_delta = np.asarray(episode_step["action_rotation_delta"], dtype=np.float64)
if robot == "google_robot_static":
# the recorded demonstration actions are in the form of axis-angle representation
gt_action_rotation_angle = np.linalg.norm(gt_action_rotation_delta)
gt_action_rotation_ax = (
gt_action_rotation_delta / gt_action_rotation_angle
if gt_action_rotation_angle > 1e-6
else np.array([0.0, 1.0, 0.0])
)
gt_action_rotation_axangle = gt_action_rotation_ax * gt_action_rotation_angle
elif robot == "widowx":
# the recorded demonstration actions are in the form of raw, pitch, yaw euler angles
gt_action_rotation_ax, gt_action_rotation_angle = euler2axangle(*gt_action_rotation_delta)
gt_action_rotation_axangle = gt_action_rotation_ax * gt_action_rotation_angle
action = np.concatenate(
[
gt_action_world_vector,
gt_action_rotation_axangle,
np.array([0]),
],
).astype(np.float64)
_ = env.step(action)
# calculate trajectory error
this_traj_err = []
this_traj_raw_transl_err = []
this_traj_raw_rot_err = []
for (tcp_pose_at_base, gt_tcp_pose_at_base) in zip(tcp_poses_at_base, gt_tcp_poses_at_base):
raw_transl_err = np.linalg.norm(tcp_pose_at_base.p - gt_tcp_pose_at_base.p)
err = raw_transl_err
this_traj_raw_transl_err.append(raw_transl_err)
R_pred = quat2mat(tcp_pose_at_base.q)
R_gt = quat2mat(gt_tcp_pose_at_base.q)
raw_rot_err = np.arcsin(
np.clip(
1 / (2 * np.sqrt(2)) * np.sqrt(np.trace((R_pred - R_gt).T @ (R_pred - R_gt))),
0.0,
1.0,
)
)
err = err + raw_rot_err
this_traj_raw_rot_err.append(raw_rot_err)
this_traj_err.append(err)
if np.mean(this_traj_err) > 0.15:
for (tcp_pose_at_base, gt_tcp_pose_at_base) in zip(tcp_poses_at_base, gt_tcp_poses_at_base):
print(tcp_pose_at_base, gt_tcp_pose_at_base)
print("*" * 10)
print(this_traj_err)
print("-" * 20)
if this_traj_err[0] > 0.02:
print(
"WARNING: The robot is not initialized to have the same pose as the first step of the episode. Error is: ",
this_traj_err[0],
)
return (
np.mean(this_traj_err),
np.mean(this_traj_raw_transl_err),
np.mean(this_traj_raw_rot_err),
)
def calc_pose_err(dset, arm_stiffness, arm_damping, robot, control_mode, log_path):
errs = []
raw_transl_errs = []
raw_rot_errs = []
processes = []
# calculate the pose error for each episode in the dataset in parallel
pool = mp.Pool(min(len(dset), 18))
for episode in dset:
processes.append(
pool.apply_async(
calc_pose_err_single_ep,
args=(episode, arm_stiffness, arm_damping, robot, control_mode),
)
)
pool.close()
for process in processes:
result = process.get()
errs.append(result[0])
raw_transl_errs.append(result[1])
raw_rot_errs.append(result[2])
pool.join()
avg_err = np.mean(errs)
avg_raw_transl_err = np.mean(raw_transl_errs)
avg_raw_rot_err = np.mean(raw_rot_errs)
# log the results
print_info = f"arm_stiffness: {list(arm_stiffness)}, arm_damping: {list(arm_damping)}, avg_raw_transl_err: {avg_raw_transl_err}, avg_raw_rot_err: {avg_raw_rot_err}, avg_err: {avg_err}, per_traj_err: {errs}"
with open(log_path, "a") as f:
print(print_info, file=f)
print(print_info)
return avg_err
if __name__ == "__main__":
"""
python tools/sysid/sysid.py --dataset-path /home/xuanlin/Downloads/sysid_dataset.pkl \
--log-path /home/xuanlin/Downloads/opt_results.txt --robot google_robot_static
python tools/sysid/sysid.py --dataset-path /home/xuanlin/Downloads/sysid_dataset_bridge.pkl \
--log-path /home/xuanlin/Downloads/opt_results_bridge.txt --robot widowx
"""
os.environ["DISPLAY"] = ""
parser = argparse.ArgumentParser()
parser.add_argument("--dataset-path", type=str, default="sysid_log/sysid_dataset.pkl")
parser.add_argument("--log-path", type=str, default="sysid_log/opt_results_google_robot.txt")
parser.add_argument("--robot", type=str, default="google_robot_static")
args = parser.parse_args()
with open(args.dataset_path, "rb") as f:
dset = pickle.load(f)
if args.robot == "google_robot_static":
control_mode = "arm_pd_ee_delta_pose_align_interpolate_by_planner_gripper_pd_joint_pos"
# these are just examples of the stiffness / damping ranges and initial values;
stiffness_high = np.array([2000, 2000, 1300, 1300, 1300, 600, 600])
stiffness_low = np.array([1400, 1400, 900, 900, 900, 380, 380])
damping_high = np.array([1200, 1200, 850, 850, 850, 350, 360])
damping_low = np.array([700, 700, 580, 580, 580, 200, 200])
init_stiffness = np.array([1700, 1700, 1100, 1000, 1100, 500, 500])
init_damping = np.array([950, 950, 700, 700, 700, 300, 300])
# stiffness_high = np.array([1800, 1750, 1050, 960, 1230, 450, 480])
# stiffness_low = np.array([1700, 1650, 950, 930, 1180, 410, 450])
# damping_high = np.array([1110, 1100, 800, 750, 730, 310, 360])
# damping_low = np.array([980, 900, 700, 610, 610, 220, 250])
# # stiffness_high = np.array([1900, 1850, 1150, 1000, 1280, 500, 530])
# # stiffness_low = np.array([1500, 1550, 850, 800, 1030, 380, 380])
# # damping_high = np.array([1200, 1200, 850, 780, 780, 350, 360])
# # damping_low = np.array([830, 830, 630, 550, 500, 190, 230])
# init_stiffness = np.array([1700.0, 1737.0471680861954, 979.975871856535, 930.0, 1212.154500274304, 432.96500923932535, 468.0013365498738])
# init_damping = np.array([1059.9791902443303, 1010.4720585373592, 767.2803161582076, 680.0, 674.9946964336588, 274.613381336198, 340.532560578637])
elif args.robot == "widowx":
control_mode = "arm_pd_ee_target_delta_pose_align2_gripper_pd_joint_pos"
stiffness_high = np.array([1400, 1400, 1400, 1400, 1400, 1400])
stiffness_low = np.array([600, 600, 600, 600, 600, 600])
damping_high = np.array([500, 400, 400, 400, 400, 400])
damping_low = np.array([150, 130, 100, 100, 100, 100])
init_stiffness = np.array([1000, 1000, 1000, 1000, 1000, 1000])
init_damping = np.array([300, 250, 200, 200, 200, 200])
# stiffness_high = np.array([1200, 780, 860, 1230, 1430, 1080])
# stiffness_low = np.array([1110, 680, 730, 1010, 1180, 930])
# damping_high = np.array([500, 350, 210, 400, 260, 330])
# damping_low = np.array([250, 150, 100, 240, 150, 200])
# init_stiffness = np.array([1169.7891719504198, 730.0, 808.4601346394447, 1229.1299089624076, 1272.2760456418862, 1056.3326605132252])
# init_damping = np.array([330.0, 180.0, 152.12036565582588, 309.6215302722146, 201.04998711007383, 269.51458932695414])
else:
raise NotImplementedError()
raw_action_to_stiffness = lambda x: stiffness_low + (stiffness_high - stiffness_low) * x[: len(stiffness_high)]
raw_action_to_damping = (
lambda x: damping_low + (damping_high - damping_low) * x[len(stiffness_high) : 2 * len(stiffness_high)]
)
init_action = np.concatenate(
[
(init_stiffness - stiffness_low) / (stiffness_high - stiffness_low),
(init_damping - damping_low) / (damping_high - damping_low),
]
)
opt_fxn = lambda x: calc_pose_err(
dset,
raw_action_to_stiffness(x),
raw_action_to_damping(x),
args.robot,
control_mode,
log_path=args.log_path,
)
opt = sa.minimize(
opt_fxn,
init_action,
opt_mode="continuous",
step_max=2000,
t_max=1.5,
t_min=0,
bounds=[[0, 1]] * (len(init_stiffness) * 2),
)
"""
# log_stiffness_high = np.log(stiffness_high)
# log_stiffness_low = np.log(stiffness_low)
# log_init_stiffness = np.log(init_stiffness)
# log_damping_high = np.log(damping_high)
# log_damping_low = np.log(damping_low)
# log_init_damping = np.log(init_damping)
# raw_action_to_stiffness = lambda x: np.exp(log_stiffness_low + (log_stiffness_high - log_stiffness_low) * x[: len(x) // 2])
# raw_action_to_damping = lambda x: np.exp(log_damping_low + (log_damping_high - log_damping_low) * x[len(x) // 2 :])
# init_action = np.concatenate(
# [(log_init_stiffness - log_stiffness_low) / (log_stiffness_high - log_stiffness_low),
# (log_init_damping - log_damping_low) / (log_damping_high - log_damping_low)]
# )
"""
|