fix: clamp std before distribution, lower init_noise to 0.5, NaN guard

exports_go1_no_linevel/metadata.txt

@@ -0,0 +1,29 @@
+# Go1 No-Linevel Terrain Walk - ONNX Policy Metadata
+env: go1-stairs-terrain-walk-no-linevel
+obs_dim: 57
+action_dim: 12
+hidden_dims: [512, 256, 128]
+
+# Observation layout (57 dims, NO linear velocity):
+  [0:3]   gyro * ang_vel_scale
+  [3:6]   gravity (body frame)
+  [6:18]  joint_angle_deviation * dof_pos_scale
+  [18:30] joint_vel * dof_vel_scale
+  [30:42] last_actions (raw)
+  [42:45] commands [vx*2.0, vy*2.0, wz*0.25]
+  [45:57] foot_contact_forces (body frame, raw)
+
+# Joint order: ['FR_hip', 'FR_thigh', 'FR_calf', 'FL_hip', 'FL_thigh', 'FL_calf', 'RR_hip', 'RR_thigh', 'RR_calf', 'RL_hip', 'RL_thigh', 'RL_calf']
+default_angles: [-0.0, 0.8999999761581421, -1.7999999523162842, 0.0, 0.8999999761581421, -1.7999999523162842, -0.0, 0.8999999761581421, -1.7999999523162842, 0.0, 0.8999999761581421, -1.7999999523162842]
+action_scale: 0.05
+kp: 80.0
+kd: 1.0
+clip_actions: 23.7
+clip_observations: 100.0
+
+# Observation scales:
+  ang_vel: 0.25
+  dof_pos: 1.0
+  dof_vel: 0.05
+  contact_force: 1.0
+  command_scale: [2.0, 2.0, 0.25]