Principle:ARISE Initiative Robomimic Checkpoint Loading

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

Overview

A checkpoint deserialization pattern that reconstructs a fully-functional rollout policy from a saved model file by chaining config recovery, observation initialization, algorithm instantiation, and weight restoration.

Description

Checkpoint Loading is the inverse operation of model saving: it takes a .pth checkpoint file and produces a ready-to-use policy for environment rollouts. The process is complex because the checkpoint must bootstrap the entire framework state: the configuration system, observation registries, algorithm class selection, network architecture construction, and weight loading.

The reconstruction chain proceeds in strict order:

1. Load the raw checkpoint dictionary from the .pth file
2. Extract the algorithm name and recreate the config object
3. Initialize observation utilities with the config
4. Extract shape metadata to determine network dimensions
5. Reconstruct normalization statistics from lists to numpy arrays
6. Instantiate the algorithm via algo_factory with correct shapes
7. Deserialize saved weights into the model
8. Wrap in a RolloutPolicy for inference

This self-contained design means any checkpoint file can be deployed without access to the original training code or configuration files.

Usage

Use this principle at the start of any evaluation or deployment pipeline. It is the first step in the Trained Policy Evaluation workflow and is also used during training when loading models for fine-tuning or transfer learning.

Theoretical Basis

# Abstract checkpoint loading chain (not real implementation)
ckpt = torch.load("model.pth")

# 1. Recover config
algo_name = ckpt["algo_name"]
config = config_factory(algo_name, dic=ckpt["config"])

# 2. Initialize observation system
initialize_obs_utils_with_config(config)

# 3. Reconstruct model architecture
model = algo_factory(algo_name, config, ckpt["shape_metadata"]["all_shapes"],
                     ckpt["shape_metadata"]["ac_dim"], device)

# 4. Load weights
model.deserialize(ckpt["model"])

# 5. Wrap for rollout
policy = RolloutPolicy(model, obs_normalization_stats=ckpt.get("obs_normalization_stats"))

Related Pages

Implemented By

• Implementation:ARISE_Initiative_Robomimic_FileUtils_policy_from_checkpoint