python:genesis-world 四足机器狗行走示例
genesis-world
下载 genesis-world-main.zip 解压缩到 D:\AI
pypi: genesis-world
pip install torch torchvision torchaudio
pip install genesis-world
genesis_world-1.4.5-py3-none-any.whl 83.6MB
运行 cmd
cd D:\AI\genesis-world-main
python examples/rigid/single_franka.py -v
RL algorithms = Reinforcement Learning algorithms,强化学习算法
一句话:
强化学习(RL)是机器学习三大分支之一(另外两个:监督学习、无监督学习);RL algorithms 就是用来实现强化学习的各类算法。
核心思想:智能体(agent)在环境里不断试错,通过拿到的奖励(reward)去学习一套动作策略,让长期累积奖励最大化。
典型用途
机器人控制、游戏 AI(AlphaGo、OpenAI Five)、自动驾驶、推荐系统、资源调度、大模型里的 RLHF(基于人类反馈的强化学习)。
说明:下面给两套脚本:
- 简易开环三角步态(直接跑,不用RL预训练权重,Unitree Go2),适合快速看行走效果;
- 官方原生RL训练/评估(
go2_train.py / go2_eval.py,和仓库原版对齐),适合做PPO训练步态
前置安装
# 先装pytorch,再装genesis-world
pip install torch torchvision
pip install genesis-world
# 如果要跑RL版本:
pip install rsl-rl-lib tensorboard
注意:genesis-world 要求 Python >=3.10
脚本1:简易开环步态示例(go2_openloop_walk.py)
原理:Trot 对角步态,正弦轨迹生成关节目标角度,PD控制,不需要训练权重,开箱可视化。
import genesis as gs
import torch
import math
def main():
gs.init(backend=gs.cpu) # 换成 gs.gpu 如果有cuda
scene = gs.Scene(
sim_options=gs.options.SimOptions(dt=0.02, substeps=2),
viewer_options=gs.options.ViewerOptions(
camera_pos=(3, 0, 2),
camera_lookat=(0,0,0.5),
camera_fov=40,
),
show_viewer=True,
)
# 地面
plane = scene.add_entity(gs.morphs.Plane())
# 加载Go2 URDF,genesis内置asset
robot = scene.add_entity(
gs.morphs.URDF(
file="urdf/go2/urdf/go2.urdf",
pos=(0, 0, 0.4),
quat=(1, 0, 0, 0),
fixed=False,
)
)
scene.build()
# 关节名称顺序对应Go2
joint_names = [
"FR_hip_joint", "FR_thigh_joint", "FR_calf_joint",
"FL_hip_joint", "FL_thigh_joint", "FL_calf_joint",
"RR_hip_joint", "RR_thigh_joint", "RR_calf_joint",
"RL_hip_joint", "RL_thigh_joint", "RL_calf_joint",
]
# 官方标准方式获取dof索引,替换掉不存在的 get_dofs_index
dofs_idx = [robot.get_joint(name).dofs_idx_local[0] for name in joint_names]
n_dof = len(dofs_idx)
# PD gains
kp = torch.tensor([40.0]*n_dof, device=scene.device)
kd = torch.tensor([2.0]*n_dof, device=scene.device)
t = 0.0
gait_freq = 2.0 # 步态频率 Hz
step_len = 0.12
height = 0.08
while True:
# Trot对角步态:FR & RL 一组,FL & RR 一组,相位差pi
phase1 = 2 * math.pi * gait_freq * t
phase2 = phase1 + math.pi
def leg_traj(phase):
swing = max(0, math.sin(phase))
hip = 0.15
thigh = -0.4 + swing * step_len
calf = 0.8 - swing * height
return [hip, thigh, calf]
q_target = []
q_target.extend(leg_traj(phase1)) # FR
q_target.extend(leg_traj(phase2)) # FL
q_target.extend(leg_traj(phase2)) # RR
q_target.extend(leg_traj(phase1)) # RL
q_target = torch.tensor(q_target, device=scene.device, dtype=torch.float32)
robot.set_dofs_target(
dofs_idx=dofs_idx,
target_pos=q_target,
kp=kp,
kd=kd,
)
scene.step()
t += 0.02
if __name__ == "__main__":
main()
运行:
python go2_openloop_walk.py
缺点:开环步态很容易摔倒,仅用于理解关节控制;稳定行走需要RL策略(下面官方版本)
脚本2:官方RL版本(Go2 PPO行走,仓库原版结构)
分成两个文件:
go2_env.py(环境) +go2_train.py(训练),训练完用eval跑
运行 cmd
cd D:\AI\genesis-world-main
python examples/locomotion/go2_env.py
python examples/locomotion/go2_backflip.py # 后空翻
python examples/locomotion/go2_train.py --max-iterations 11
python examples/locomotion/go2_eval.py -e go2-walking --ckpt 10
go2_env.py
import torch
import genesis as gs
class Go2Env:
def __init__(self, num_envs=1, device="cpu"):
self.num_envs = num_envs
self.device = device
self.dt = 0.02
self.substeps = 2
self.scene = gs.Scene(
sim_options=gs.options.SimOptions(dt=self.dt, substeps=self.substeps),
rigid_options=gs.options.RigidOptions(enable_self_collision=False),
viewer_options=gs.options.ViewerOptions(),
show_viewer=(num_envs == 1),
)
self.scene.add_entity(gs.morphs.Plane())
self.robot = self.scene.add_entity(
gs.morphs.URDF(file="urdf/go2/urdf/go2.urdf", pos=(0,0,0.4))
)
self.scene.build(n_envs=num_envs)
self.joint_names = [
"FR_hip_joint", "FR_thigh_joint", "FR_calf_joint",
"FL_hip_joint", "FL_thigh_joint", "FL_calf_joint",
"RR_hip_joint", "RR_thigh_joint", "RR_calf_joint",
"RL_hip_joint", "RL_thigh_joint", "RL_calf_joint",
]
# 官方标准方式获取dof索引,替换掉不存在的 get_dofs_index
self.dofs_idx = [self.robot.get_joint(name).dofs_idx_local[0] for name in self.joint_names]
self.n_dof = len(self.dofs_idx)
self.default_joint_pos = torch.tensor([
0.1, 0.8, -1.5,
-0.1, 0.8, -1.5,
0.1, 0.8, -1.5,
-0.1, 0.8, -1.5,
], device=device)
self.kp = torch.full((num_envs, self.n_dof), 40.0, device=device)
self.kd = torch.full((num_envs, self.n_dof), 2.0, device=device)
self.commands = torch.zeros((num_envs,3), device=device) # vx, vy, yaw
self.obs_buf = torch.zeros((num_envs, 48), device=device)
self.reset_buf = torch.ones(num_envs, device=device, dtype=torch.bool)
def reset_idx(self, env_ids):
self.robot.set_dofs_pos(
self.dofs_idx,
self.default_joint_pos[None, :].repeat(len(env_ids),1),
env_indices=env_ids
)
self.robot.set_dofs_vel(self.dofs_idx, torch.zeros(len(env_ids), self.n_dof, device=self.device), env_indices=env_ids)
self.commands[env_ids, 0] = 0.6 # 前向速度指令
self.reset_buf[env_ids] = False
def step(self, actions):
actions = torch.clip(actions, -1, 1) * 0.25
target_pos = self.default_joint_pos[None, :] + actions
self.robot.set_dofs_target(
self.dofs_idx, target_pos,
kp=self.kp, kd=self.kd
)
self.scene.step()
base_pos = self.robot.get_base_pos()
base_lin_vel = self.robot.get_base_lin_vel()
base_ang_vel = self.robot.get_base_ang_vel()
dof_pos = self.robot.get_dofs_pos(self.dofs_idx)
dof_vel = self.robot.get_dofs_vel(self.dofs_idx)
self.obs_buf = torch.cat([
base_lin_vel, base_ang_vel, self.commands,
dof_pos - self.default_joint_pos, dof_vel
], dim=-1)
reward = self._compute_reward(base_lin_vel)
termination = (base_pos[:,2] < 0.25)
self.reset_buf |= termination
self.reset_idx(torch.where(self.reset_buf)[0])
return self.obs_buf, reward, self.reset_buf, {}
def _compute_reward(self, base_lin_vel):
err = torch.square(self.commands[:,0] - base_lin_vel[:,0])
r = torch.exp(-err / 0.25)
return r
go2_train.py
import torch
from rsl_rl.runners import OnPolicyRunner
from go2_env import Go2Env
def main():
device = "cpu"
env = Go2Env(num_envs=4096, device=device)
cfg = {
"policy": {"init_noise_std": 1.0, "hidden_dims": [512,256,128]},
"runner": {"max_iterations": 1000},
}
runner = OnPolicyRunner(env, cfg, log_dir="logs/go2_walk", device=device)
runner.learn()
if __name__ == "__main__":
main()
训练:
python go2_train.py
训练后加载权重eval(你可以在eval脚本里打开viewer,单env看行走)
官方原版仓库:https://github.com/Genesis-Embodied-AI/genesis-world/tree/main/examples/locomotion
常见问题
- 找不到
urdf/go2/urdf/go2.urdf:genesis内置asset,新版会自动拉取;如果缺失,从官方仓库assets目录复制 - 开环步态容易倒:属于正常现象,开环没有状态反馈;稳定必须RL
- 切换GPU:
gs.init(backend=gs.gpu)
如果你想要,可以再给你一个训练完成后,加载ckpt单独可视化行走的eval脚本。
下面是配套的 eval 可视化脚本,加载训练好的 ckpt,单环境、打开 viewer,直接看机器狗行走。
注意:和上面
go2_env.py保持在同一个目录,权重路径改成你训练输出的路径。
# go2_eval.py
import torch
import genesis as gs
from rsl_rl.runners import OnPolicyRunner
from go2_env import Go2Env
def main():
# -------------------------- 配置区 --------------------------
ckpt_path = "logs/go2_walk/model_1000.pt" # 修改成你的ckpt路径
device = "cpu" # 有cuda就改成 "cuda:0"
# 指令:vx, vy, yaw(前进速度、横向速度、转向角速度)
cmd_vx = 0.6
cmd_vy = 0.0
cmd_yaw = 0.0
# -----------------------------------------------------------
gs.init(backend=gs.cpu if device=="cpu" else gs.gpu)
# 评估只用 1 个环境,开启可视化
env = Go2Env(num_envs=1, device=device)
# 覆盖指令,让机器狗往前走
env.commands[0, 0] = cmd_vx
env.commands[0, 1] = cmd_vy
env.commands[0, 2] = cmd_yaw
# 加载runner和policy
cfg = {
"policy": {"init_noise_std": 1.0, "hidden_dims": [512,256,128]},
"runner": {"max_iterations": 1000},
}
runner = OnPolicyRunner(env, cfg, log_dir="logs/go2_walk", device=device)
runner.load(ckpt_path)
policy = runner.get_inference_policy(device=device)
policy.eval()
obs, _ = env.reset_idx(torch.tensor([0], device=device))
obs = env.obs_buf.clone()
with torch.no_grad():
while True:
actions = policy(obs)
obs, reward, done, info = env.step(actions)
# 如果摔倒重置
if done.any():
env.reset_idx(torch.where(done)[0])
obs = env.obs_buf.clone()
if __name__ == "__main__":
main()
使用方法
- 先跑训练脚本
go2_train.py,训练结束后,logs/go2_walk/下会生成model_xxxx.pt - 修改
ckpt_path指向你对应的权重文件 - 运行:
python go2_eval.py
几个可改的地方
- 想让机器狗转弯:把
cmd_yaw改成正数/负数(比如0.8) - 想原地转向:
cmd_vx=0, cmd_yaw=0.8 - 换成GPU:
device="cuda:0",同时gs.init(backend=gs.gpu) - 如果训练时 policy 网络结构改了,eval 里的
cfg["policy"]必须和训练完全一致,否则加载失败
重要提醒
- rsl-rl 加载权重时,训练和评估的观测维度、动作维度、网络结构必须严格一致
- 如果训练时加了噪声、动作缩放、奖励函数改动,eval 的环境类
go2_env.py不能和训练版本不一样 - 若出现
KeyError权重加载失败,大概率是 cfg 不匹配
更多推荐
所有评论(0)