Model management

Every saved model is a self-describing directory. See artifacts for the format.

Save

policy is an rfx.nn.Policy instance (e.g., MLP, ActorCritic) after training. robot_config and normalizer are optional — pass them to make the artifact self-describing:

from rfx.nn import MLP
from rfx.utils import ObservationNormalizer

policy = MLP(obs_dim=48, action_dim=12)     # or any trained Policy
config = rfx.GO2_CONFIG                      # RobotConfig (built-in or custom)
normalizer = ObservationNormalizer()         # optional, bundles obs stats

# ...train policy + fit normalizer...

policy.save(
    "runs/go2-walk-v1",
    robot_config=config,
    normalizer=normalizer,
    training_info={"total_steps": 50000},
)

See Policies for how to define and train a policy. Creates:

runs/go2-walk-v1/
├── rfx_config.json     # architecture + robot + training metadata
├── model.safetensors   # weights
└── normalizer.json     # observation normalizer state

Load

loaded = rfx.load_policy("runs/go2-walk-v1")
loaded = rfx.load_policy("hf://rfx-community/go2-walk-v1")

loaded.policy         # reconstructed policy
loaded.robot_config   # RobotConfig or None
loaded.normalizer     # normalizer or None
loaded.policy_type    # "MLP", "ActorCritic", ...

LoadedPolicy is callable and handles torch / tinygrad conversion automatically.

Inspect

config = rfx.inspect_policy("runs/go2-walk-v1")
print(config["policy_type"])     # "MLP"
print(config["policy_config"])   # {"obs_dim": 48, ...}

HuggingFace Hub

Push:

rfx.push_policy("runs/go2-walk-v1", "rfx-community/go2-walk-v1")

Pull (at load time):

loaded = rfx.load_policy("hf://rfx-community/go2-walk-v1")

rfx deploy calls load_policy internally. If the artifact’s rfx_config.json records a robot, --robot is inferred automatically.

​Save

​Load

​Inspect

​HuggingFace Hub

Save

Load

Inspect

HuggingFace Hub