change rewards
This commit is contained in:
@@ -14,41 +14,42 @@ from rl_game.get_up.env.t1_env import T1SceneCfg
|
||||
import isaaclab.envs.mdp as mdp
|
||||
|
||||
|
||||
# --- 1. 自定义 MDP 逻辑函数 (放在配置类之前) ---
|
||||
# --- 1. 修正后的自定义 MDP 逻辑函数 ---
|
||||
|
||||
def is_standing_still(
|
||||
env: ManagerBasedRLEnv,
|
||||
minimum_height: float,
|
||||
max_angle_error: float,
|
||||
standing_time: float
|
||||
standing_time: float,
|
||||
velocity_threshold: float = 0.2
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
判定机器人是否稳定站立了一段时间。
|
||||
逻辑：高度达标且姿态垂直，持续时间超过 standing_time 则返回 True。
|
||||
判定判定逻辑：高度达标 + 躯干垂直 + 几乎静止 + 维持时间。
|
||||
增加了速度判定，彻底杜绝起跳瞬间触发。
|
||||
"""
|
||||
# 获取当前状态：高度 (Z轴) 和 投影重力 (前两个分量越小越垂直)
|
||||
|
||||
# 1. 首先获取 H2 (头部) 的索引 (建议在环境初始化时获取一次，或者如下所示获取)
|
||||
# find_bodies 返回 (indices, names)
|
||||
# 1. 获取 Body 索引
|
||||
head_idx, _ = env.scene["robot"].find_bodies("H2")
|
||||
|
||||
# 2. 修改后的位置获取逻辑
|
||||
# data.body_link_pos_w 的维度是 (num_envs, num_bodies, 3)
|
||||
# 我们取所有环境 (:)，对应的头部索引 (head_idx[0])，以及 Z 轴坐标 (2)
|
||||
current_head_height = env.scene["robot"].data.body_link_pos_w[:, head_idx[0], 2]
|
||||
|
||||
# 3. 姿态判定保持不变（通常依然以躯干 Trunk 的垂直度为准，因为头部可能会摆动）
|
||||
# 2. 状态量获取
|
||||
current_head_height = env.scene["robot"].data.body_state_w[:, head_idx[0], 2]
|
||||
# 投影重力误差（越小越垂直）
|
||||
gravity_error = torch.norm(env.scene["robot"].data.projected_gravity_b[:, :2], dim=-1)
|
||||
# 根部线速度和角速度（判定是否晃动）
|
||||
root_vel_norm = torch.norm(env.scene["robot"].data.root_lin_vel_w, dim=-1)
|
||||
root_ang_vel_norm = torch.norm(env.scene["robot"].data.root_ang_vel_w, dim=-1)
|
||||
|
||||
# 4. 更新判断逻辑
|
||||
is_stable_now = (current_head_height > minimum_height) & (gravity_error < max_angle_error)
|
||||
# 3. 综合判定（这里不强制检查力，改用更稳健的速度限制）
|
||||
is_stable_now = (
|
||||
(current_head_height > minimum_height) &
|
||||
(gravity_error < max_angle_error) &
|
||||
(root_vel_norm < velocity_threshold) &
|
||||
(root_ang_vel_norm < velocity_threshold * 2.0)
|
||||
)
|
||||
|
||||
# 在 env.extras 中维护一个计时器
|
||||
# 4. 计时器
|
||||
if "stable_timer" not in env.extras:
|
||||
env.extras["stable_timer"] = torch.zeros(env.num_envs, device=env.device)
|
||||
|
||||
# 核心计时逻辑：达标累加 dt，不达标清零
|
||||
# dt = physics_dt * decimation (即控制频率的步长)
|
||||
dt = env.physics_dt * env.cfg.decimation
|
||||
env.extras["stable_timer"] = torch.where(
|
||||
is_stable_now,
|
||||
@@ -56,60 +57,53 @@ def is_standing_still(
|
||||
torch.zeros_like(env.extras["stable_timer"])
|
||||
)
|
||||
|
||||
# 判定是否达到目标时长
|
||||
return env.extras["stable_timer"] > standing_time
|
||||
|
||||
|
||||
def get_success_reward(env: ManagerBasedRLEnv, term_keys: str) -> torch.Tensor:
|
||||
"""检查是否触发了特定的成功终止条件"""
|
||||
return env.termination_manager.get_term(term_keys)
|
||||
|
||||
|
||||
def root_vel_z_l2_local(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
|
||||
"""专门惩罚 Z 轴方向的速度"""
|
||||
# 获取根部速度 (num_envs, 3) -> [vx, vy, vz]
|
||||
vel = env.scene[asset_cfg.name].data.root_lin_vel_w
|
||||
# 只取 Z 轴：vel[:, 2]
|
||||
return torch.square(vel[:, 2])
|
||||
# 严厉惩罚 Z 轴正向速度（向上窜）
|
||||
vel_z = env.scene[asset_cfg.name].data.root_lin_vel_w[:, 2]
|
||||
return torch.square(torch.clamp(vel_z, min=0.0))
|
||||
|
||||
def joint_torques_l2_local(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
|
||||
"""计算机器人所有关节施加扭矩的平方和"""
|
||||
# 从 data.applied_torques 获取数据，通常形状为 (num_envs, num_joints)
|
||||
torques = env.scene[asset_cfg.name].data.applied_torque
|
||||
return torch.sum(torch.square(torques), dim=-1)
|
||||
|
||||
def joint_vel_l2_local(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
|
||||
"""计算机器人所有关节速度的平方和"""
|
||||
# 从 data.joint_vel 获取数据
|
||||
vel = env.scene[asset_cfg.name].data.joint_vel
|
||||
return torch.sum(torch.square(vel), dim=-1)
|
||||
def feet_airtime_penalty_local(
|
||||
env: ManagerBasedRLEnv,
|
||||
sensor_cfg: SceneEntityCfg,
|
||||
threshold: float = 1.0
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
自定义滞空惩罚逻辑：
|
||||
如果脚部的垂直合力小于阈值，说明脚离地了。
|
||||
返回一个 Tensor，离地时为 1.0，着地时为 0.0。
|
||||
"""
|
||||
# 1. 获取传感器对象
|
||||
contact_sensor = env.scene.sensors.get(sensor_cfg.name)
|
||||
|
||||
def joint_pos_limits_l2_local(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
|
||||
"""惩罚关节接近或超过限位"""
|
||||
# 获取关节位置 (num_envs, num_joints)
|
||||
joint_pos = env.scene[asset_cfg.name].data.joint_pos
|
||||
# 获取限位 (num_joints, 2) -> [lower, upper]
|
||||
limits = env.scene[asset_cfg.name].data.soft_joint_pos_limits
|
||||
lower_limits = limits[..., 0]
|
||||
upper_limits = limits[..., 1]
|
||||
if contact_sensor is None:
|
||||
# 如果没搜到传感器，返回全 0，防止程序崩溃
|
||||
return torch.zeros(env.num_envs, device=env.device)
|
||||
|
||||
# 计算超出限位的部分
|
||||
out_of_lower = torch.clamp(lower_limits - joint_pos, min=0.0)
|
||||
out_of_upper = torch.clamp(joint_pos - upper_limits, min=0.0)
|
||||
# 2. 获取触地力 (num_envs, num_bodies_in_sensor, 3)
|
||||
# 我们取所有被监测 Body (左右脚) 的 Z 轴推力
|
||||
# 如果所有脚的力都小于 threshold，判定为“完全腾空”
|
||||
foot_forces_z = contact_sensor.data.net_forces_w[:, :, 2]
|
||||
is_in_air = torch.all(foot_forces_z < threshold, dim=-1)
|
||||
|
||||
# 返回超出量的平方和
|
||||
return torch.sum(torch.square(out_of_lower + out_of_upper), dim=-1)
|
||||
return is_in_air.float()
|
||||
|
||||
# --- 2. 配置类定义 ---
|
||||
|
||||
## 1. 定义与你的类一致的关节列表 (按照 ROBOT_MOTORS 的顺序)
|
||||
T1_JOINT_NAMES = [
|
||||
'Left_Hip_Pitch', 'Right_Hip_Pitch',
|
||||
'Left_Hip_Roll', 'Right_Hip_Roll',
|
||||
'Left_Hip_Yaw', 'Right_Hip_Yaw',
|
||||
'Left_Knee_Pitch', 'Right_Knee_Pitch',
|
||||
'Left_Ankle_Pitch', 'Right_Ankle_Pitch',
|
||||
'Left_Ankle_Roll', 'Right_Ankle_Roll'
|
||||
'Left_Hip_Pitch', 'Right_Hip_Pitch', 'Left_Hip_Roll', 'Right_Hip_Roll',
|
||||
'Left_Hip_Yaw', 'Right_Hip_Yaw', 'Left_Knee_Pitch', 'Right_Knee_Pitch',
|
||||
'Left_Ankle_Pitch', 'Right_Ankle_Pitch', 'Left_Ankle_Roll', 'Right_Ankle_Roll'
|
||||
]
|
||||
|
||||
|
||||
@configclass
|
||||
class T1ObservationCfg:
|
||||
"""观察值空间配置：严格对应你的 Robot 基类数据结构"""
|
||||
@@ -129,21 +123,10 @@ class T1ObservationCfg:
|
||||
|
||||
# 3. 重力投影 (对应 global_orientation_euler/quat 相关的姿态感知)
|
||||
projected_gravity = ObsTerm(func=mdp.projected_gravity)
|
||||
|
||||
# 4. 关节位置 (对应 motor_positions)
|
||||
# 使用 joint_pos_rel 获取相对于默认姿态的偏差，显式指定关节顺序
|
||||
joint_pos = ObsTerm(
|
||||
func=mdp.joint_pos_rel,
|
||||
params={"asset_cfg": SceneEntityCfg("robot", joint_names=T1_JOINT_NAMES)}
|
||||
)
|
||||
|
||||
# 5. 关节速度 (对应 motor_speeds)
|
||||
joint_vel = ObsTerm(
|
||||
func=mdp.joint_vel_rel,
|
||||
params={"asset_cfg": SceneEntityCfg("robot", joint_names=T1_JOINT_NAMES)}
|
||||
)
|
||||
|
||||
# 6. 上一次的动作 (对应 motor_targets)
|
||||
joint_pos = ObsTerm(func=mdp.joint_pos_rel,
|
||||
params={"asset_cfg": SceneEntityCfg("robot", joint_names=T1_JOINT_NAMES)})
|
||||
joint_vel = ObsTerm(func=mdp.joint_vel_rel,
|
||||
params={"asset_cfg": SceneEntityCfg("robot", joint_names=T1_JOINT_NAMES)})
|
||||
actions = ObsTerm(func=mdp.last_action)
|
||||
|
||||
policy = PolicyCfg()
|
||||
@@ -158,7 +141,7 @@ class T1EventCfg:
|
||||
"asset_cfg": SceneEntityCfg("robot"),
|
||||
"pose_range": {
|
||||
"roll": (0, 0),#(-1.57, 1.57),
|
||||
"pitch": (1.57, 1.57),#(-1.57, 1.57),
|
||||
"pitch": (-1.57, 1.57),#(-1.57, 1.57),
|
||||
"yaw": (0, 0),#(-3.14, 3.14),
|
||||
"x": (0.0, 0.0),
|
||||
"y": (0.0, 0.0),
|
||||
@@ -172,124 +155,89 @@ class T1EventCfg:
|
||||
|
||||
@configclass
|
||||
class T1ActionCfg:
|
||||
"""动作空间"""
|
||||
"""关键修改：降低 scale 让动作变丝滑，增大阻尼效果"""
|
||||
joint_pos = JointPositionActionCfg(
|
||||
asset_name="robot",
|
||||
joint_names=T1_JOINT_NAMES,
|
||||
scale=0.5,
|
||||
scale=0.2, # 从 0.5 降到 0.2，防止电机暴力抽搐
|
||||
use_default_offset=True
|
||||
)
|
||||
|
||||
|
||||
@configclass
|
||||
class T1GetUpRewardCfg:
|
||||
"""优化后的奖励函数：抑制跳跃，引导稳健起身"""
|
||||
# 1. 姿态奖 (核心引导)
|
||||
upright = RewTerm(func=mdp.flat_orientation_l2, weight=5.0)
|
||||
|
||||
# 1. 高度引导 (改为平滑的指数奖励)
|
||||
# 相比 root_height_below_minimum，这个函数会让机器人越接近目标高度得分越高，且曲线平稳
|
||||
# 2. 只有脚着地时才给的高度奖（模拟逻辑）
|
||||
# 修正：直接使用 root_height 配合强力的速度惩罚
|
||||
height_tracking = RewTerm(
|
||||
func=mdp.root_height_below_minimum, # 如果没有自定义函数，保留这个但调低权重
|
||||
weight=2.0, # 降低权重，防止“弹射”
|
||||
func=mdp.root_height_below_minimum,
|
||||
weight=2.0,
|
||||
params={
|
||||
"minimum_height": 1.05,
|
||||
"asset_cfg": SceneEntityCfg("robot", body_names="H2"),
|
||||
}
|
||||
)
|
||||
|
||||
# 2. 姿态奖 (保持不变，这是核心)
|
||||
upright = RewTerm(func=mdp.flat_orientation_l2, weight=5.0)
|
||||
|
||||
# 3. 稳定性引导 (增加对速度的惩罚，抑制跳跃)
|
||||
# 惩罚过大的垂直速度，防止“跳起”
|
||||
# 3. 抑制跳跃：严厉惩罚向上窜的速度
|
||||
root_vel_z_penalty = RewTerm(
|
||||
func=root_vel_z_l2_local, # 使用本地函数
|
||||
weight=-5,
|
||||
params={"asset_cfg": SceneEntityCfg("robot")} # 传入资产配置
|
||||
func=root_vel_z_l2_local,
|
||||
weight=-15.0, # 增大负权重
|
||||
params={"asset_cfg": SceneEntityCfg("robot")}
|
||||
)
|
||||
|
||||
feet_contact = RewTerm(
|
||||
func=mdp.contact_forces,
|
||||
weight=0.5,
|
||||
# 4. 抑制滞空 (Airtime Penalty)
|
||||
# 如果脚离开地面，按时间扣分
|
||||
feet_airtime = RewTerm(
|
||||
func=feet_airtime_penalty_local,
|
||||
weight=-15.0,
|
||||
params={
|
||||
"sensor_cfg": SceneEntityCfg("feet_contact_sensor"),
|
||||
"threshold": 1.0
|
||||
"threshold": 0.2, # 超过 0.2s 离地就开始扣分
|
||||
}
|
||||
)
|
||||
|
||||
# 4. 关节与能量约束 (防止 NaN 和乱跳的关键)
|
||||
joint_vel = RewTerm(
|
||||
func=joint_vel_l2_local,
|
||||
weight=-1,
|
||||
params={"asset_cfg": SceneEntityCfg("robot")}
|
||||
)
|
||||
# 5. 动作平滑 (非常重要：解决视频中的高频抖动)
|
||||
action_rate = RewTerm(func=mdp.action_rate_l2, weight=-1.0)
|
||||
joint_vel = RewTerm(func=mdp.joint_vel_l2, weight=-0.1, params={"asset_cfg": SceneEntityCfg("robot")})
|
||||
|
||||
applied_torque = RewTerm(
|
||||
func=joint_torques_l2_local,
|
||||
weight=-1.0e-2,
|
||||
params={"asset_cfg": SceneEntityCfg("robot")}
|
||||
)
|
||||
# 6. 时间惩罚：逼迫它快点站稳
|
||||
time_penalty = RewTerm(func=mdp.is_alive, weight=-0.5)
|
||||
|
||||
# 5. 动作平滑 (非常重要)
|
||||
action_rate = RewTerm(
|
||||
func=mdp.action_rate_l2,
|
||||
weight=-1.0 # 增大权重，强制动作连贯
|
||||
)
|
||||
|
||||
# 6. 软限位惩罚：防止关节撞击
|
||||
joint_limits = RewTerm(
|
||||
func=joint_pos_limits_l2_local,
|
||||
weight=-10.0,
|
||||
params = {"asset_cfg": SceneEntityCfg("robot")}
|
||||
)
|
||||
|
||||
# 7. 时间惩罚 (强制效率)
|
||||
# 每一帧都扣除固定分数，迫使机器人尽快达成 is_success 以停止扣分
|
||||
time_penalty = RewTerm(
|
||||
func=mdp.is_alive,
|
||||
weight=-1.2
|
||||
)
|
||||
|
||||
# 8. 核心终点奖励
|
||||
# 7. 终极奖励
|
||||
is_success = RewTerm(
|
||||
func=get_success_reward,
|
||||
weight=500.0, # 成功奖励可以给高点，但前提是动作要平稳
|
||||
weight=200.0, # 成功一次给大奖，但判定条件极严
|
||||
params={"term_keys": "standing_success"}
|
||||
)
|
||||
|
||||
@configclass
|
||||
class T1GetUpTerminationsCfg:
|
||||
"""终止条件：站稳即算任务完成，且包含强制超时重置"""
|
||||
|
||||
# --- 关键：必须显式添加这一行，episode_length_s 才会生效 ---
|
||||
time_out = DoneTerm(func=mdp.time_out)
|
||||
|
||||
# 失败：跌倒 (Trunk 倾斜过大)
|
||||
# limit_angle 是弧度，1.0 约等于 57度，如果想严格点可以调小
|
||||
base_crash = DoneTerm(func=mdp.bad_orientation, params={"limit_angle": 1.0})
|
||||
# 严格的失败判定：躯干倾斜超过 45 度就重置，不让它在地上滚
|
||||
base_crash = DoneTerm(func=mdp.bad_orientation, params={"limit_angle": 0.785})
|
||||
|
||||
# 成功：满足“头部高度”和“姿态要求”，且维持 0.8 秒
|
||||
# 严格的成功判定
|
||||
standing_success = DoneTerm(
|
||||
func=is_standing_still, # 确保你已经把这个函数里的 current_height 改成了 H2 的 Z 轴
|
||||
func=is_standing_still,
|
||||
params={
|
||||
# T1 头部 (H2) 站直高度约 1.15-1.2m，设为 1.1m 比较稳健
|
||||
"minimum_height": 1.05,
|
||||
# 姿态误差 (投影重力分量)，0.15 约等于 8.6 度，要求很直
|
||||
"max_angle_error": 1.0,
|
||||
# 维持时间
|
||||
"standing_time": 0.8
|
||||
"minimum_height": 1.05, # H2 高度
|
||||
"max_angle_error": 0.15, # 极小角度误差（约 8.6 度）
|
||||
"standing_time": 0.8, # 必须保持 0.8 秒（起跳不可能在空中停这么久）
|
||||
"velocity_threshold": 0.15 # 速度必须极低
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
@configclass
|
||||
class T1EnvCfg(ManagerBasedRLEnvCfg):
|
||||
"""主环境配置"""
|
||||
scene = T1SceneCfg(num_envs=16384, env_spacing=2.5)
|
||||
scene = T1SceneCfg(num_envs=16384, env_spacing=2.5) # 建议先用 4096 验证逻辑
|
||||
|
||||
def __post_init__(self):
|
||||
super().__post_init__()
|
||||
# 初始高度设低，配合随机旋转事件实现“从地上爬起来”
|
||||
self.scene.robot.init_state.pos = (0.0, 0.0, 0.4)
|
||||
self.scene.robot.init_state.pos = (0.0, 0.0, 0.4) # 初始稍微抬高一点点，防止卡地
|
||||
|
||||
observations = T1ObservationCfg()
|
||||
rewards = T1GetUpRewardCfg()
|
||||
@@ -297,5 +245,5 @@ class T1EnvCfg(ManagerBasedRLEnvCfg):
|
||||
events = T1EventCfg()
|
||||
actions = T1ActionCfg()
|
||||
|
||||
episode_length_s = 10.0 # 3秒强制重置
|
||||
decimation = 4 # 控制频率
|
||||
episode_length_s = 5.0 # 缩短时长，增加训练效率
|
||||
decimation = 4
|
||||
Reference in New Issue
Block a user