Apollo3D/Gym_GPU
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

change reward function

Browse Source
This commit is contained in:
ChenXi
2026-03-21 07:00:49 -04:00
parent 6d2ad9846a
commit d78fdeda0d
1 changed files with 27 additions and 147 deletions
Download Patch File Download Diff File Expand all files Collapse all files

174
rl_game/get_up/config/t1_env_cfg.py
View File

@@ -23,46 +23,30 @@ def standing_with_feet_reward(
min_head_height: float,
min_pelvis_height: float,
sensor_cfg: SceneEntityCfg,
force_threshold: float = 30.0,
max_v_z: float = 0.25
force_threshold: float = 20.0,
max_v_z: float = 0.5
) -> torch.Tensor:
"""
平滑切换的高度奖励：
低高度 -> 纯高度引导
高高度 -> 高度 + 足底力 + 速度约束
"""
# 1. 获取基本状态
head_idx, _ = env.scene["robot"].find_bodies("H2")
pelvis_idx, _ = env.scene["robot"].find_bodies("Trunk")
current_head_h = env.scene["robot"].data.body_state_w[:, head_idx[0], 2]
current_pelvis_h = env.scene["robot"].data.body_state_w[:, pelvis_idx[0], 2]
# 2. 计算基础高度得分 (0.0 - 1.0)
head_score = torch.clamp(current_head_h / min_head_height, max=1.0)
pelvis_score = torch.clamp(current_pelvis_h / min_pelvis_height, max=1.0)
combined_height_score = (head_score + pelvis_score) / 2.0
curr_head_h = torch.clamp(env.scene["robot"].data.body_state_w[:, head_idx[0], 2], 0.0, 2.0)
curr_pelvis_h = torch.clamp(env.scene["robot"].data.body_state_w[:, pelvis_idx[0], 2], 0.0, 2.0)
head_score = torch.tanh(curr_head_h / (min_head_height + 1e-6) * 2.0)
pelvis_score = torch.tanh(curr_pelvis_h / (min_pelvis_height + 1e-6) * 2.0)
height_reward = (head_score + pelvis_score) / 2.0
# 3. 计算足底力判定
contact_sensor = env.scene.sensors.get(sensor_cfg.name)
foot_forces_z = torch.sum(contact_sensor.data.net_forces_w[:, :, 2], dim=-1)
is_feet_on_ground = (foot_forces_z > force_threshold).float()
# 4. 计算速度惩罚 (抑制乱跳)
force_weight = torch.sigmoid((foot_forces_z - force_threshold) / 5.0)
root_vel_z = env.scene["robot"].data.root_lin_vel_w[:, 2]
vel_penalty_factor = torch.exp(-4.0 * torch.clamp(torch.abs(root_vel_z) - max_v_z, min=0.0))
vel_penalty = torch.exp(-2.0 * torch.clamp(torch.abs(root_vel_z) - max_v_z, min=0.0))
# --- 核心逻辑切换 ---
# 定义一个“过渡高度” (例如盆骨达到 0.4m)
transition_h = 0.4
# 如果高度很低：给纯高度奖，诱导它向上动
low_height_reward = combined_height_score
# 如果高度较高：给 综合奖 (高度 * 速度限制 * 必须踩地)
high_height_reward = combined_height_score * vel_penalty_factor * is_feet_on_ground
return torch.where(current_pelvis_h < transition_h, low_height_reward, high_height_reward)
influence_weight = torch.clamp((curr_pelvis_h - 0.2) / 0.4, min=0.0, max=1.0)
combined_reward = height_reward * ((1.0 - influence_weight) + influence_weight * force_weight * vel_penalty)
return combined_reward
def arm_push_up_reward(
env: ManagerBasedRLEnv,
@@ -105,18 +89,6 @@ def arm_push_up_reward(
pushing_up_bonus,
torch.zeros_like(pushing_up_bonus))
def linear_head_height_reward(env: ManagerBasedRLEnv, target_height: float, base_height: float = 0.15) -> torch.Tensor:
"""
计算头部从地面到目标高度的线性增量奖励
"""
head_idx, _ = env.scene["robot"].find_bodies("H2")
current_head_h = env.scene["robot"].data.body_state_w[:, head_idx[0], 2]
# 计算相对于地面的提升量，并归一化到 0-1
reward = (current_head_h - base_height) / (target_height - base_height)
return torch.clamp(reward, min=0.0, max=1.0)
def is_standing_still(
env: ManagerBasedRLEnv,
min_head_height: float,
@@ -152,51 +124,6 @@ def is_standing_still(
return env.extras["stable_timer"] > standing_time
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)
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)
# 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)
return is_in_air.float()
def root_vel_z_l2_local(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
# 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_pos_rel_l2_local(env: ManagerBasedRLEnv, asset_cfg: SceneEntityCfg) -> torch.Tensor:
# 获取相对默认位置的偏差 (num_envs, num_joints)
rel_pos = mdp.joint_pos_rel(env, asset_cfg)
# 计算平方和 (L2)
return torch.sum(torch.square(rel_pos), dim=-1)
def strict_feet_contact_reward(env: ManagerBasedRLEnv, sensor_cfg: SceneEntityCfg) -> torch.Tensor:
"""如果脚不着地，直接给一个很大的负分，强制它必须寻找支点"""
contact_sensor = env.scene.sensors.get(sensor_cfg.name)
# 只要有一只脚没力，就判定为不稳
foot_forces_z = contact_sensor.data.net_forces_w[:, :, 2]
all_feet_cond = torch.min(foot_forces_z, dim=-1)[0] > 5.0 # 左右脚都要有至少5N的力
return (~all_feet_cond).float() # 返回1表示违规
# --- 2. 配置类 ---
T1_JOINT_NAMES = [
@@ -263,19 +190,20 @@ class T1GetUpRewardCfg:
# 2. 【条件高度奖】：双高度判定（头+盆骨），且必须脚踩地
height_with_feet = RewTerm(
func=standing_with_feet_reward,
weight=50.0, # 作为核心引导，增加权重
weight=25.0, # 作为核心引导，增加权重
params={
"min_head_height": 1.10,
"min_pelvis_height": 0.65,
"min_pelvis_height": 0.7,
"sensor_cfg": SceneEntityCfg("contact_sensor", body_names=[".*_foot_link"]),
"force_threshold": 10.0
"force_threshold": 20.0,
"max_v_z": 0.3
}
)
# 3. 手臂撑地奖：辅助脱离地面阶段
arm_push_support = RewTerm(
func=arm_push_up_reward,
weight=20.0, # 显著增加权重（从 3.0 提到 15.0），让它成为起步的关键
weight=15.0, # 显著增加权重（从 3.0 提到 15.0），让它成为起步的关键
params={
"sensor_cfg": SceneEntityCfg("contact_sensor", body_names=[".*_hand_link", "AL3", "AR3"]),
"height_threshold": 0.6, # 躯干升到 0.6m 前都鼓励手臂用力
@@ -283,66 +211,18 @@ class T1GetUpRewardCfg:
}
)
# 4. 引导机器人“向上看”和“抬起头”
head_lift = RewTerm(
func=linear_head_height_reward,
weight=15.0,
params={"target_height": 1.1, "base_height": 0.15}
)
# 5. 惩罚项
undesired_contacts = RewTerm(
func=mdp.undesired_contacts,
weight=-2.0,
params={
"sensor_cfg": SceneEntityCfg("contact_sensor", body_names=["H2", "Trunk"]),
# 注意：此处必须排除手臂相关 link，否则手臂用力时会同时被扣分
"threshold": 1.0
}
)
# 6. 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,
weight=-1.0, # 增大负权重
# 4. 关节限位惩罚 (新增：防止关节撞死导致数值问题)
joint_limits = RewTerm(
func=mdp.joint_pos_limits,
weight=-1.0,
params={"asset_cfg": SceneEntityCfg("robot")}
)
# 7. 抑制滞空 (Airtime Penalty)
feet_airtime = RewTerm(
func=strict_feet_contact_reward,
weight=-5.0, # 加大权重，跳一下扣的分比站起来得的分还多
params={"sensor_cfg": SceneEntityCfg("contact_sensor", body_names=[".*_foot_link"])}
)
joint_vel_penalty = RewTerm(
func=mdp.joint_vel_l2,
weight=-0.01, # 惩罚过快的关节运动
params={"asset_cfg": SceneEntityCfg("robot")}
)
action_rate = RewTerm(
func=mdp.action_rate_l2,
weight=-0.05, # 惩罚动作的突变，让动作更丝滑，减少爆发力
)
# 惩罚躯干的翻转和俯仰角速度
base_ang_vel_penalty = RewTerm(
func=lambda env, asset_cfg: torch.norm(mdp.base_ang_vel(env, asset_cfg), dim=-1),
weight=-0.1,
params={"asset_cfg": SceneEntityCfg("robot")}
)
joint_deviation = RewTerm(
func=joint_pos_rel_l2_local,
weight=0.1, # 权重不要太高，只是为了让它动起来
params={"asset_cfg": SceneEntityCfg("robot")}
)
# 7. 成功终极大奖
# 5. 成功终极大奖
is_success = RewTerm(
func=lambda env, keys: env.termination_manager.get_term(keys),
weight=300.0,
func=lambda env, keys: env.termination_manager.get_term(keys).float(),
weight=50.0,
params={"keys": "standing_success"}
)
@@ -382,4 +262,4 @@ class T1EnvCfg(ManagerBasedRLEnvCfg):
actions = T1ActionCfg()
episode_length_s = 6.0
decimation = 2
decimation = 4