decrease extreme explorationin training

This commit is contained in:
2026-08-27 21:55:16 +02:00
parent 4cc2d37d94
commit 7b9c52955b
+3 -3
View File
@@ -87,7 +87,7 @@ def main():
args.pretrained_model, args.pretrained_model,
env=vec_env, env=vec_env,
learning_rate=5e-5, # Lower learning rate so RL fine-tunes without destroying base gait learning_rate=5e-5, # Lower learning rate so RL fine-tunes without destroying base gait
ent_coef=0.001, ent_coef=0.0,
target_kl=0.05, target_kl=0.05,
vf_coef=0.5, vf_coef=0.5,
max_grad_norm=0.5, max_grad_norm=0.5,
@@ -107,7 +107,7 @@ def main():
gamma=0.99, gamma=0.99,
gae_lambda=0.95, gae_lambda=0.95,
clip_range=0.2, clip_range=0.2,
ent_coef=0.001, ent_coef=0.0,
target_kl=0.05, target_kl=0.05,
vf_coef=0.5, vf_coef=0.5,
max_grad_norm=0.5, max_grad_norm=0.5,
@@ -115,7 +115,7 @@ def main():
tensorboard_log=args.log_dir, tensorboard_log=args.log_dir,
device="cpu", device="cpu",
) )
model.policy.log_std.data.fill_(-2.0)
# Setup Callbacks with ppo<number> naming # Setup Callbacks with ppo<number> naming
checkpoint_callback = CheckpointCallback( checkpoint_callback = CheckpointCallback(
save_freq=max(1, args.save_freq // args.num_workers), save_freq=max(1, args.save_freq // args.num_workers),