SinjaeKang/wp_robomimic

WarmPrior Robomimic Checkpoints

This repository contains model checkpoints from the WarmPrior_Robomimic experiments.

Repository Structure

Each experiment run is stored in its own folder with the following structure:

run_name/
├── model_best.pt   # Best performing checkpoint
└── model_latest.pt # Latest checkpoint

Available Runs

Run Name	Method	Best	Latest
PA_0116_010604_tool_hang_ph_image_flow_chitransformer_384_noise0.5_seed2	PA	✅	✅
PA_0116_010619_tool_hang_ph_image_flow_chiunet_256_noise0.5_seed2	PA	✅	✅
PA_0116_010624_square_mh_image_flow_chitransformer_384_noise0.5_seed2	PA	✅	✅
PA_0116_010629_square_mh_image_flow_chiunet_256_noise0.5_seed2	PA	✅	✅
PA_0117_141516_square_mh_image_flow_chiunet_256_noise0.3_seed2	PA	✅	✅
PA_0117_165708_square_mh_image_flow_chitransformer_384_noise0.3_seed2	PA	✅	✅
PA_0117_211126_tool_hang_ph_image_flow_chiunet_256_noise0.3_seed2	PA	✅	✅
PA_0117_231750_tool_hang_ph_image_flow_chitransformer_384_noise0.3_seed2	PA	✅	✅

Usage

To download a specific checkpoint:

from huggingface_hub import hf_hub_download
import torch

# Download a specific checkpoint
model_path = hf_hub_download(
    repo_id="SinjaeKang/wp_robomimic",
    filename="RUN_NAME/model_best.pt"
)

# Load the model
checkpoint = torch.load(model_path, map_location='cpu')

Or download all checkpoints:

from huggingface_hub import snapshot_download

# Download entire repository
local_dir = snapshot_download(
    repo_id="SinjaeKang/wp_robomimic",
    local_dir="./checkpoints"
)

Methods

• BASE: Baseline model without warm prior
• PA: Previous Action warm prior
• AC: Action Chunk warm prior

Training Details

For more information about the training procedure and methodology, please refer to the WarmPrior_Robomimic repository.