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

exports_go1_flat/metadata.txt

@@ -0,0 +1,21 @@
+# Go1 Flat Terrain Walk - ONNX Policy Metadata
+obs_dim: 45
+action_dim: 12
+hidden_dims: [256, 128, 64]
+observation_layout: gyro(3) + gravity(3) + joint_angle(12) + joint_vel(12) + last_action(12) + command(3)
+  - NO linear velocity in observation
+
+# 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 (applied BEFORE ONNX normalization):
+  lin_vel: 2.0
+  ang_vel: 0.25
+  dof_pos: 1.0
+  dof_vel: 0.05
+  command_scale: [2.0, 2.0, 0.25]  # for [vx, vy, wz]