Apollo3D/Gym_CPU
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

improve train speed and add speed constrain

Browse Source
This commit is contained in:
xxh
2026-03-13 08:51:49 -04:00
parent 3a42120857
commit 6ab356a947
4 changed files with 317 additions and 281 deletions
Download Patch File Download Diff File Expand all files Collapse all files

271
scripts/gyms/Walk.py
View File

@@ -1,11 +1,11 @@
import os
import numpy as np
import math
from time import sleep
import time
from time import sleep
from random import random
from random import uniform
from stable_baselines3 import PPO
from stable_baselines3.common.vec_env import SubprocVecEnv
@@ -28,25 +28,24 @@ Learn how to run forward using step primitive
- class Train: implements algorithms to train a new model or test an existing model
'''
class WalkEnv(gym.Env):
def __init__(self, ip, server_p) -> None:
# Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
self.Player = player = Base_Agent(
team_name="Gym",
number=1,
host=ip,
port=server_p
team_name="Gym",
number=1,
host=ip,
port=server_p
)
self.robot_type = self.Player.robot
self.step_counter = 0 # to limit episode size
self.robot_type = self.Player.robot
self.step_counter = 0 # to limit episode size
self.force_play_on = True
self.target_position = np.array([0.0, 0.0]) # target position in the x-y plane
self.initial_position = np.array([0.0, 0.0]) # initial position in the x-y plane
self.target_direction = 0.0 # target direction in the x-y plane (relative to the robot's orientation)
self.target_position = np.array([0.0, 0.0]) # target position in the x-y plane
self.initial_position = np.array([0.0, 0.0]) # initial position in the x-y plane
self.target_direction = 0.0 # target direction in the x-y plane (relative to the robot's orientation)
self.isfallen = False
self.waypoint_index = 0
self.route_completed = False
@@ -55,16 +54,26 @@ class WalkEnv(gym.Env):
self.auto_calibrate_train_sim_flip = True
self.nominal_calibrated_once = False
self.flip_calibrated_once = False
self._target_hz = 0.0
self._target_dt = 0.0
self._last_sync_time = None
target_hz_env = 24
if target_hz_env:
try:
self._target_hz = float(target_hz_env)
except ValueError:
self._target_hz = 0.0
if self._target_hz > 0.0:
self._target_dt = 1.0 / self._target_hz
# State space
# 原始观测大小: 78
obs_size = 78
self.obs = np.zeros(obs_size, np.float32)
self.observation_space = spaces.Box(
low=-10.0,
high=10.0,
shape=(obs_size,),
low=-10.0,
high=10.0,
shape=(obs_size,),
dtype=np.float32
)
@@ -73,11 +82,10 @@ class WalkEnv(gym.Env):
self.action_space = spaces.Box(
low=-1.0,
high=1.0,
shape=(action_dim,),
shape=(action_dim,),
dtype=np.float32
)
# 中立姿态
self.joint_nominal_position = np.array(
[
@@ -110,22 +118,22 @@ class WalkEnv(gym.Env):
self.train_sim_flip = np.array(
[
1.0, # 0: Head_yaw (he1)
1.0, # 0: Head_yaw (he1)
-1.0, # 1: Head_pitch (he2)
1.0, # 2: Left_Shoulder_Pitch (lae1)
1.0, # 2: Left_Shoulder_Pitch (lae1)
-1.0, # 3: Left_Shoulder_Roll (lae2)
1.0, # 4: Left_Elbow_Pitch (lae3)
1.0, # 5: Left_Elbow_Yaw (lae4)
1.0, # 4: Left_Elbow_Pitch (lae3)
1.0, # 5: Left_Elbow_Yaw (lae4)
-1.0, # 6: Right_Shoulder_Pitch (rae1)
1.0, # 7: Right_Shoulder_Roll (rae2)
1.0, # 8: Right_Elbow_Pitch (rae3)
1.0, # 9: Right_Elbow_Yaw (rae4)
1.0, # 10: Waist (te1)
1.0, # 11: Left_Hip_Pitch (lle1)
1.0, # 7: Right_Shoulder_Roll (rae2)
1.0, # 8: Right_Elbow_Pitch (rae3)
1.0, # 9: Right_Elbow_Yaw (rae4)
1.0, # 10: Waist (te1)
1.0, # 11: Left_Hip_Pitch (lle1)
-1.0, # 12: Left_Hip_Roll (lle2)
-1.0, # 13: Left_Hip_Yaw (lle3)
1.0, # 14: Left_Knee_Pitch (lle4)
1.0, # 15: Left_Ankle_Pitch (lle5)
1.0, # 14: Left_Knee_Pitch (lle4)
1.0, # 15: Left_Ankle_Pitch (lle5)
-1.0, # 16: Left_Ankle_Roll (lle6)
-1.0, # 17: Right_Hip_Pitch (rle1)
-1.0, # 18: Right_Hip_Roll (rle2)
@@ -141,7 +149,7 @@ class WalkEnv(gym.Env):
self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
self.previous_pos = np.array([0.0, 0.0]) # Track previous position
self.Player.server.connect()
sleep(2.0) # Longer wait for connection to establish completely
# sleep(2.0) # Longer wait for connection to establish completely
self.Player.server.send_immediate(
f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
)
@@ -160,7 +168,7 @@ class WalkEnv(gym.Env):
changed = []
for idx, (reference_value, observed_value) in enumerate(
zip(self.reference_joint_nominal_position, neutral_joint_positions)
zip(self.reference_joint_nominal_position, neutral_joint_positions)
):
if idx >= 10:
continue
@@ -187,38 +195,37 @@ class WalkEnv(gym.Env):
height = float(self.Player.world.global_position[2])
reliable = (
leg_norm > 0.8
and leg_max > 0.35
and 0.12 < height < 0.8
leg_norm > 0.8
and leg_max > 0.35
and 0.12 < height < 0.8
)
return reliable, leg_norm, leg_max, height
def observe(self, init=False):
"""获取当前观测值"""
robot = self.Player.robot
world = self.Player.world
# Safety check: ensure data is available
# 计算目标速度
raw_target = self.target_position - world.global_position[:2]
velocity = MathOps.rotate_2d_vec(
raw_target,
-robot.global_orientation_euler[2],
raw_target,
-robot.global_orientation_euler[2],
is_rad=False
)
# 计算相对方向
rel_orientation = MathOps.vector_angle(velocity) * 0.3
rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
velocity = np.concatenate([velocity, np.array([rel_orientation])])
velocity[0] = np.clip(velocity[0], -0.5, 0.5)
velocity[1] = np.clip(velocity[1], -0.25, 0.25)
# 关节状态
radian_joint_positions = np.deg2rad(
[robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
@@ -226,21 +233,20 @@ class WalkEnv(gym.Env):
radian_joint_speeds = np.deg2rad(
[robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
)
qpos_qvel_previous_action = np.concatenate([
(radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
radian_joint_speeds / 110.0 * self.train_sim_flip,
self.previous_action / 10.0,
])
# 角速度
ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
# 投影的重力方向
orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
# 组合观测
observation = np.concatenate([
qpos_qvel_previous_action,
@@ -249,8 +255,6 @@ class WalkEnv(gym.Env):
projected_gravity,
])
observation = np.clip(observation, -10.0, 10.0)
return observation.astype(np.float32)
@@ -260,6 +264,17 @@ class WalkEnv(gym.Env):
self.Player.world.update()
self.Player.robot.commit_motor_targets_pd()
self.Player.server.send()
if self._target_dt > 0.0:
now = time.time()
if self._last_sync_time is None:
self._last_sync_time = now
return
elapsed = now - self._last_sync_time
remaining = self._target_dt - elapsed
if remaining > 0.0:
time.sleep(remaining)
now = time.time()
self._last_sync_time = now
def debug_joint_status(self):
robot = self.Player.robot
@@ -295,12 +310,11 @@ class WalkEnv(gym.Env):
if seed is not None:
np.random.seed(seed)
length1 = np.random.uniform(10, 20) # randomize target distance
length2 = np.random.uniform(10, 20) # randomize target distance
length3 = np.random.uniform(10, 20) # randomize target distance
angle2 = np.random.uniform(-30, 30) # randomize initial orientation
angle3 = np.random.uniform(-30, 30) # randomize target direction
length1 = np.random.uniform(10, 20) # randomize target distance
length2 = np.random.uniform(10, 20) # randomize target distance
length3 = np.random.uniform(10, 20) # randomize target distance
angle2 = np.random.uniform(-30, 30) # randomize initial orientation
angle3 = np.random.uniform(-30, 30) # randomize target direction
self.step_counter = 0
self.waypoint_index = 0
@@ -322,12 +336,12 @@ class WalkEnv(gym.Env):
# 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
finished_count = 0
for _ in range(20):
for _ in range(10):
finished = self.Player.skills_manager.execute("Neutral")
self.sync()
if finished:
finished_count += 1
if finished_count >= 2: # 假设需要连续2次完成才算成功
if finished_count >= 3: # 假设需要连续3次完成才算成功
break
# neutral_joint_positions = np.deg2rad(
@@ -356,17 +370,15 @@ class WalkEnv(gym.Env):
# reset_action_noise = np.random.uniform(-0.015, 0.015, size=(len(self.Player.robot.ROBOT_MOTORS),))
# self.target_joint_positions = (self.joint_nominal_position + reset_action_noise) * self.train_sim_flip
# for idx, target in enumerate(self.target_joint_positions):
# r.set_motor_target_position(
# r.ROBOT_MOTORS[idx], target*180/math.pi, kp=25, kd=0.6
# )
# memory variables
self.initial_position = np.array(self.Player.world.global_position[:2])
self.previous_pos = self.initial_position.copy() # Critical: set to actual position
self.act = np.zeros(self.no_of_actions ,np.float32)
self.act = np.zeros(self.no_of_actions, np.float32)
point1 = self.initial_position + np.array([length1, 0])
point2 = point1 + MathOps.rotate_2d_vec(np.array([length2, 0]), angle2, is_rad=False)
point3 = point2 + MathOps.rotate_2d_vec(np.array([length3, 0]), angle3, is_rad=False)
@@ -423,57 +435,51 @@ class WalkEnv(gym.Env):
tilt_penalty = -0.2 * np.linalg.norm(self.Player.robot.gyroscope[:2]) / 100.0
return (
progress_reward
+ speed_reward
+ direction_reward
+ alive_bonus
+ height_reward
+ motionless_penalty
+ waypoint_bonus
+ action_penalty
+ tilt_penalty
progress_reward
+ speed_reward
+ direction_reward
+ alive_bonus
+ height_reward
+ motionless_penalty
+ waypoint_bonus
+ action_penalty
+ tilt_penalty
)
def step(self, action):
r = self.Player.robot
r = self.Player.robot
self.previous_action = action
self.target_joint_positions = (
self.joint_nominal_position
+ self.scaling_factor * action
self.joint_nominal_position
+ self.scaling_factor * action
)
self.target_joint_positions *= self.train_sim_flip
for idx, target in enumerate(self.target_joint_positions):
r.set_motor_target_position(
r.ROBOT_MOTORS[idx], target*180/math.pi, kp=25, kd=0.6
r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
)
self.sync() # run simulation step
self.sync() # run simulation step
self.step_counter += 1
# if self.step_counter % self.debug_every_n_steps == 0:
# self.debug_joint_status()
current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
# Compute reward based on movement from previous step
reward = self.compute_reward(self.previous_pos, current_pos, action)
# Update previous position
self.previous_pos = current_pos.copy()
# Fall detection and penalty
is_fallen = self.Player.world.global_position[2] < 0.3
# terminal state: the robot is falling or timeout
terminated = is_fallen or self.step_counter > 800 or self.route_completed
truncated = False
@@ -481,64 +487,65 @@ class WalkEnv(gym.Env):
return self.observe(), reward, terminated, truncated, {}
class Train(Train_Base):
def __init__(self, script) -> None:
super().__init__(script)
def train(self, args):
#--------------------------------------- Learning parameters
# --------------------------------------- Learning parameters
n_envs = 8 # Reduced from 8 to decrease CPU/network pressure during init
if n_envs < 1:
raise ValueError("GYM_CPU_N_ENVS must be >= 1")
n_steps_per_env = 512 # RolloutBuffer is of size (n_steps_per_env * n_envs)
minibatch_size = 64 # should be a factor of (n_steps_per_env * n_envs)
minibatch_size = 128 # should be a factor of (n_steps_per_env * n_envs)
total_steps = 30000000
learning_rate = 3e-4
folder_name = f'Walk_R{self.robot_type}'
model_path = f'./scripts/gyms/logs/{folder_name}/'
model_path = f'./scripts/gyms/logs/{folder_name}/'
print(f"Model path: {model_path}")
print(f"Using {n_envs} parallel environments")
#--------------------------------------- Run algorithm
# --------------------------------------- Run algorithm
def init_env(i_env):
def thunk():
return WalkEnv( self.ip , self.server_p + i_env)
return WalkEnv(self.ip, self.server_p + i_env)
return thunk
servers = Train_Server( self.server_p, self.monitor_p_1000, n_envs+1 ) #include 1 extra server for testing
server_log_dir = os.path.join(model_path, "server_logs")
os.makedirs(server_log_dir, exist_ok=True)
servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1) # include 1 extra server for testing
# Wait for servers to start
print(f"Starting {n_envs+1} rcssservermj servers...")
print(f"Starting {n_envs + 1} rcssservermj servers...")
print("Servers started, creating environments...")
env = SubprocVecEnv( [init_env(i) for i in range(n_envs)] )
eval_env = SubprocVecEnv( [init_env(n_envs)] )
env = SubprocVecEnv([init_env(i) for i in range(n_envs)])
eval_env = SubprocVecEnv([init_env(n_envs)])
try:
# Custom policy network architecture
policy_kwargs = dict(
net_arch=dict(
pi=[256, 256, 128], # Policy network: 3 layers
vf=[256, 256, 128] # Value network: 3 layers
vf=[256, 256, 128] # Value network: 3 layers
),
activation_fn=__import__('torch.nn', fromlist=['ReLU']).ReLU,
)
if "model_file" in args: # retrain
model = PPO.load( args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env, batch_size=minibatch_size, learning_rate=learning_rate )
else: # train new model
model = PPO(
"MlpPolicy",
env=env,
verbose=1,
n_steps=n_steps_per_env,
batch_size=minibatch_size,
learning_rate=learning_rate,
if "model_file" in args: # retrain
model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
batch_size=minibatch_size, learning_rate=learning_rate)
else: # train new model
model = PPO(
"MlpPolicy",
env=env,
verbose=1,
n_steps=n_steps_per_env,
batch_size=minibatch_size,
learning_rate=learning_rate,
device="cpu",
policy_kwargs=policy_kwargs,
ent_coef=0.01, # Entropy coefficient for exploration
@@ -547,28 +554,32 @@ class Train(Train_Base):
gamma=0.99 # Discount factor
)
model_path = self.learn_model( model, total_steps, model_path, eval_env=eval_env, eval_freq=n_steps_per_env*20, save_freq=n_steps_per_env*20, backup_env_file=__file__ )
model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
eval_freq=n_steps_per_env * 10, save_freq=n_steps_per_env * 10,
backup_env_file=__file__)
except KeyboardInterrupt:
sleep(1) # wait for child processes
sleep(1) # wait for child processes
print("\nctrl+c pressed, aborting...\n")
servers.kill()
return
env.close()
eval_env.close()
servers.kill()
def test(self, args):
# Uses different server and monitor ports
server = Train_Server( self.server_p-1, self.monitor_p, 1 )
env = WalkEnv( self.ip, self.server_p-1 )
model = PPO.load( args["model_file"], env=env )
server_log_dir = os.path.join(args["folder_dir"], "server_logs")
os.makedirs(server_log_dir, exist_ok=True)
server = Train_Server(self.server_p - 1, self.monitor_p, 1, log_dir=server_log_dir)
env = WalkEnv(self.ip, self.server_p - 1)
model = PPO.load(args["model_file"], env=env)
try:
self.export_model( args["model_file"], args["model_file"]+".pkl", False ) # Export to pkl to create custom behavior
self.test_model( model, env, log_path=args["folder_dir"], model_path=args["folder_dir"] )
self.export_model(args["model_file"], args["model_file"] + ".pkl",
False) # Export to pkl to create custom behavior
self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
except KeyboardInterrupt:
print()
@@ -578,18 +589,18 @@ class Train(Train_Base):
if __name__ == "__main__":
from types import SimpleNamespace
# 创建默认参数
script_args = SimpleNamespace(
args=SimpleNamespace(
i='127.0.0.1', # Server IP
p=3100, # Server port
m=3200, # Monitor port
r=0, # Robot type
t='Gym', # Team name
u=1 # Uniform number
p=3100, # Server port
m=3200, # Monitor port
r=0, # Robot type
t='Gym', # Team name
u=1 # Uniform number
)
)
trainer = Train(script_args)
trainer.train({})
trainer.train({"model_file": "scripts/gyms/logs/Walk_R0_000/model_245760_steps.zip"})