File size: 7,738 Bytes
c399dd1 | 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 | """Train new PPO goal-reaching policies on the ROM (Sim2Sim campaign, Case 2).
Copies the release GoalEnv dynamics, reward and PPO hyperparameters (rl_script/rl_train_goal_reach.py)
into a Gymnasium env, because SB3 2.7 needs Gymnasium. Only the training-noise config and the seed change.
Configs:
nonoise : no action noise; powertrain fixed at tau0=1000, omega0=2600.
release : release semantics. N(0,0.05) on throttle and steering each step; tau0, omega0 drawn once
at env construction with sd 50 and 100.
robust : N(0,0.10) action noise; tau0, omega0 redrawn every episode with sd 100 and 200;
actuation delay drawn every episode uniform from {0,1,2} control steps (0-0.2 s).
Usage: python train_ppo.py --config release --seed 11 --out policies/release_s11
"""
import argparse
import json
import platform
import time
from pathlib import Path
import gymnasium as gym
import numpy as np
import torch as th
from gymnasium import spaces
from stable_baselines3 import PPO
from stable_baselines3.common.callbacks import BaseCallback
from stable_baselines3.common.vec_env import DummyVecEnv
CONFIGS = {
'nonoise': dict(act_sigma=0.0, tau_sigma=0.0, omega_sigma=0.0, powertrain_redraw='never', delay_steps=[0]),
'release': dict(act_sigma=0.05, tau_sigma=50.0, omega_sigma=100.0, powertrain_redraw='construction', delay_steps=[0]),
'robust': dict(act_sigma=0.10, tau_sigma=100.0, omega_sigma=200.0, powertrain_redraw='episode', delay_steps=[0, 1, 2]),
}
class GoalEnvG(gym.Env):
"""Release GoalEnv (same constants, dynamics, reward, termination), Gymnasium API, configurable noise."""
def __init__(self, cfg, seed):
super().__init__()
self.cfg = cfg
self.rng = np.random.default_rng(seed)
self.action_space = spaces.Box(low=np.array([0.2, -1.0]), high=np.array([0.5, 1.0]), shape=(2,), dtype=np.float32)
self.observation_space = spaces.Box(low=-np.inf, high=np.inf, shape=(4,), dtype=np.float32)
self.r_wheel, self.i_wheel, self.gamma = 0.47, 6.69, 1 / 3
self.c1, self.c0, self.car_length = 1e-4, 0.02, 1.0
self.dt, self.max_steps = 0.1, 400
self._draw_powertrain()
self.old_distance = 0.0
def _draw_powertrain(self):
self.tau_0 = 1000 + self.rng.normal(0, self.cfg['tau_sigma'])
self.omega_0 = 2600 + self.rng.normal(0, self.cfg['omega_sigma'])
def f_vdot(self, alpha, v):
omega_m = v / (self.r_wheel * self.gamma)
f1_m = -self.tau_0 * omega_m / self.omega_0 + self.tau_0
T_fun = alpha * f1_m - self.c1 * omega_m - self.c0
return T_fun * self.gamma / self.i_wheel * self.r_wheel
def relative_state_to_goal(self):
x, y, theta, v = self.state[:4]
rot_mat = np.array([[np.cos(theta), np.sin(theta)], [-np.sin(theta), np.cos(theta)]])
relative_pos = np.dot(rot_mat, np.array(self.goal) - np.array([x, y]))
relative_theta = np.arctan2(relative_pos[1], relative_pos[0])
return np.array([relative_pos[0], relative_pos[1], relative_theta, v], dtype=np.float32)
def reset(self, seed=None, options=None):
if self.cfg['powertrain_redraw'] == 'episode':
self._draw_powertrain()
self.delay = int(self.rng.choice(self.cfg['delay_steps']))
self.queue = [np.array([0.0, 0.0])] * self.delay
r = self.rng.uniform(10.0, 20.0)
theta = self.rng.uniform(0, 2 * np.pi)
self.goal = [r * np.cos(theta), r * np.sin(theta)]
self.state = np.array([0, 0, 0, 0.0], dtype=np.float32)
self.steps = 0
self.old_distance = np.sqrt(self.goal[0] ** 2 + self.goal[1] ** 2)
return self.relative_state_to_goal(), {}
def step(self, action):
self.queue.append(np.asarray(action, dtype=np.float64))
alpha, beta = self.queue.pop(0)
if self.cfg['act_sigma'] > 0:
alpha = alpha + self.rng.normal(0, self.cfg['act_sigma'])
beta = beta + self.rng.normal(0, self.cfg['act_sigma'])
x, y, theta, v = self.state
x_dot = np.cos(theta) * v
y_dot = np.sin(theta) * v
theta_dot = v * np.tan(beta * 0.7) / self.car_length
v_dot = self.f_vdot(alpha, v)
x += x_dot * self.dt
y += y_dot * self.dt
theta += theta_dot * self.dt
v += v_dot * self.dt
v = np.clip(v, 0.0, 5.0)
self.state = np.array([x, y, theta, v], dtype=np.float32)
self.steps += 1
distance = np.sqrt((x - self.goal[0]) ** 2 + (y - self.goal[1]) ** 2)
progress = self.old_distance - distance
reward = 0.0
if progress > 0:
reward = progress * 50
if progress < 0.1:
reward -= 100
done = False
if distance < 1.0:
reward += self.max_steps * 15
done = True
if self.steps >= self.max_steps:
reward -= 1000 * distance
done = True
self.old_distance = distance
# Release treats both goal and timeout as terminal (old Gym done=True); keep that.
return self.relative_state_to_goal(), float(reward), done, False, {'velocity': float(v)}
class SaveEvery(BaseCallback):
def __init__(self, every, out):
super().__init__()
self.every, self.out, self.k = every, Path(out), 0
self.ep_returns, self.cur = [], 0.0
def _on_step(self):
self.cur += float(self.locals['rewards'][0])
if self.locals['dones'][0]:
self.ep_returns.append(self.cur)
self.cur = 0.0
if self.n_calls % self.every == 0:
self.k += 1
self.model.save(self.out / f'ckpt_{self.k}.zip')
return True
def main():
p = argparse.ArgumentParser()
p.add_argument('--config', choices=list(CONFIGS), required=True)
p.add_argument('--seed', type=int, required=True)
p.add_argument('--out', required=True)
p.add_argument('--timesteps', type=int, default=2_000_000)
p.add_argument('--save-every', type=int, default=500_000)
a = p.parse_args()
th.set_num_threads(1)
out = Path(a.out)
out.mkdir(parents=True, exist_ok=False)
cfg = CONFIGS[a.config]
env = GoalEnvG(cfg, seed=a.seed + 7)
vec = DummyVecEnv([lambda: env])
model = PPO('MlpPolicy', vec, learning_rate=3e-4, n_steps=env.max_steps * 3, batch_size=64, n_epochs=20,
gamma=0.99, clip_range=0.2, verbose=0, seed=a.seed, device='cpu',
policy_kwargs=dict(net_arch=dict(pi=[64, 128, 128, 64], vf=[64, 128, 128, 64]), activation_fn=th.nn.ReLU))
cb = SaveEvery(a.save_every, out)
t0 = time.time()
model.learn(total_timesteps=a.timesteps, callback=cb)
rets = np.array(cb.ep_returns)
meta = dict(config=a.config, config_values=cfg, seed=a.seed, env_seed=a.seed + 7, timesteps=a.timesteps,
save_every=a.save_every, wall_s=time.time() - t0,
fixed_construction_powertrain=dict(tau0=env.tau_0, omega0=env.omega_0) if cfg['powertrain_redraw'] == 'construction' else None,
episodes=len(rets), mean_return_last100=float(rets[-100:].mean()) if len(rets) else None,
hyperparameters=dict(learning_rate=3e-4, n_steps=1200, batch_size=64, n_epochs=20, gamma=0.99, clip_range=0.2,
net_arch=[64, 128, 128, 64], activation='ReLU'),
versions=dict(python=platform.python_version(), torch=th.__version__, numpy=np.__version__,
sb3=__import__('stable_baselines3').__version__))
np.save(out / 'episode_returns.npy', rets.astype(np.float32))
(out / 'train_meta.json').write_text(json.dumps(meta, indent=2) + '\n')
if __name__ == '__main__':
main()
|