Implementation:ARISE Initiative Robomimic FileUtils policy from checkpoint

Knowledge Sources	robomimic robomimic Pretrained Models
Domains	Robotics, Evaluation, Serialization
Last Updated	2026-02-15 08:00 GMT

Overview

Concrete tool for restoring a trained policy from a checkpoint file into a ready-to-use RolloutPolicy wrapper provided by the robomimic file utilities module.

Description

The policy_from_checkpoint function implements the full checkpoint restoration chain: load checkpoint dict, recover algorithm name and config, initialize observation utils, reconstruct the model architecture via algo_factory, deserialize weights, apply normalization stats, and wrap as RolloutPolicy. It accepts either a file path or a pre-loaded dictionary to avoid redundant disk reads.

Usage

Call this function at the start of evaluation scripts (run_trained_agent.py) or when deploying trained policies. The returned RolloutPolicy can be directly called with observations to produce actions.

Code Reference

Source Location

• Repository: robomimic
• File: robomimic/utils/file_utils.py
• Lines: L373-444

Signature

def policy_from_checkpoint(device=None, ckpt_path=None, ckpt_dict=None, verbose=False):
    """
    Restores a trained policy from a checkpoint file or loaded model dictionary.

    Args:
        device (torch.device): if provided, put model on this device
        ckpt_path (str): Path to checkpoint file. Only needed if not providing @ckpt_dict.
        ckpt_dict (dict): Loaded model checkpoint dictionary. Only needed if not providing @ckpt_path.
        verbose (bool): if True, include print statements

    Returns:
        model (RolloutPolicy): policy wrapper ready for rollouts
        ckpt_dict (dict): loaded checkpoint dictionary
    """

Import

import robomimic.utils.file_utils as FileUtils

# Call as:
policy, ckpt_dict = FileUtils.policy_from_checkpoint(ckpt_path="model.pth")

I/O Contract

Inputs

Name	Type	Required	Description
device	torch.device	No	Compute device. Default: auto-detected from config
ckpt_path	str	No	Path to .pth checkpoint file (provide this or ckpt_dict)
ckpt_dict	dict	No	Pre-loaded checkpoint dictionary (provide this or ckpt_path)
verbose	bool	No	Print loading info. Default: False

Outputs

Name	Type	Description
model	RolloutPolicy	Ready-to-use policy wrapper with loaded weights, observation normalization, and action normalization
ckpt_dict	dict	Loaded checkpoint dictionary for reuse (e.g., by env_from_checkpoint)

Usage Examples

Load and Evaluate Policy

import robomimic.utils.file_utils as FileUtils
import torch

# Load policy from checkpoint
policy, ckpt_dict = FileUtils.policy_from_checkpoint(
    ckpt_path="/path/to/model_epoch_100.pth",
    device=torch.device("cuda:0"),
    verbose=True,
)

# Use policy in rollout
obs = env.reset()
policy.start_episode()
action = policy(ob=obs)

Reuse Checkpoint Dict

import robomimic.utils.file_utils as FileUtils

# Load policy
policy, ckpt_dict = FileUtils.policy_from_checkpoint(ckpt_path="model.pth")

# Reuse ckpt_dict for environment creation (avoids re-loading from disk)
env, _ = FileUtils.env_from_checkpoint(ckpt_dict=ckpt_dict, render_offscreen=True)

Related Pages

Implements Principle

• Principle:ARISE_Initiative_Robomimic_Checkpoint_Loading

Requires Environment

• Environment:ARISE_Initiative_Robomimic_PyTorch_CUDA_Environment