✨ clean up lot of stuff

configs/training/ppo_real.yaml

@@ -8,15 +8,17 @@ defaults:
   - _self_
 
 hidden_sizes: [256, 256]
-total_timesteps: 100000
-learning_epochs: 5
-learning_rate: 0.001            # conservative — can't undo real-world damage
-entropy_loss_scale: 0.0001
-log_interval: 1024
+total_timesteps: 2000000
+learning_epochs: 10
+learning_rate: 0.0005            # conservative — can't undo real-world damage
+entropy_loss_scale: 0.01
+rollout_steps: 2048
+mini_batches: 8
+log_interval: 2048
 checkpoint_interval: 5000       # frequent saves — can't rewind real hardware
 initial_log_std: -0.5           # moderate initial exploration
 min_log_std: -4.0
-max_log_std: 0.0                # cap σ at 1.0
+max_log_std: 2.0                # cap σ at 1.0
 
 # Never run real-hardware training remotely
 remote: false

configs/training/ppo_single.yaml

@@ -8,7 +8,7 @@ defaults:
   - _self_
 
 hidden_sizes: [256, 256]
-total_timesteps: 1000000
+total_timesteps: 2000000
 learning_epochs: 10
 learning_rate: 0.0003
 entropy_loss_scale: 0.01