cyy/Motrixlab
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
8x54zj-m
2026-06-30 15:35:58 +08:00
parent f2a8e0e2ff
commit b0f1da4596
61 changed files with 6522 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
scripts/play.py
View File

@@ -36,6 +36,7 @@ _RAND_SEED = flags.DEFINE_bool("rand-seed", False, "Generate random seed")
_RLLIB = flags.DEFINE_string(
"rllib", None, "The RL framework (skrl/rslrl). Auto-discovered from latest training if not specified."
)
_FORCE_PHASE = flags.DEFINE_integer("force-phase", None, "Lock terrain phase (0=flat,1=rough,2=stairs,3=mixed)")
def get_inference_backend(policy_path: Path | str, rllib: str):
@@ -179,13 +180,21 @@ def main(argv):
backend = get_inference_backend(policy_path, rllib)
# Build env config overrides
env_cfg_override = {}
if _FORCE_PHASE.present:
env_cfg_override["force_phase"] = _FORCE_PHASE.value
if not env_cfg_override:
env_cfg_override = None
if rllib == "rslrl":
# RSLRL evaluation flow (always uses torch backend)
assert device_supports.torch, "PyTorch is not available on your device"
from motrix_rl.rslrl.torch.train import ppo
config.torch.backend = "torch"
trainer = ppo.Trainer(env_name, sim_backend, cfg_override=rl_override, enable_render=enable_render)
trainer = ppo.Trainer(env_name, sim_backend, cfg_override=rl_override,
enable_render=enable_render, env_cfg_override=env_cfg_override)
trainer.play(policy_path)
elif backend == "jax":
@@ -193,7 +202,8 @@ def main(argv):
from motrix_rl.skrl.jax.train import ppo
config.jax.backend = "jax" # or "numpy"
trainer = ppo.Trainer(env_name, sim_backend, cfg_override=rl_override, enable_render=enable_render)
trainer = ppo.Trainer(env_name, sim_backend, cfg_override=rl_override,
enable_render=enable_render, env_cfg_override=env_cfg_override)
trainer.play(policy_path)
elif backend == "torch":
@@ -201,7 +211,8 @@ def main(argv):
from motrix_rl.skrl.torch.train import ppo
config.torch.backend = "torch"
trainer = ppo.Trainer(env_name, sim_backend, cfg_override=rl_override, enable_render=enable_render)
trainer = ppo.Trainer(env_name, sim_backend, cfg_override=rl_override,
enable_render=enable_render, env_cfg_override=env_cfg_override)
trainer.play(policy_path)