You can deploy this model and start earning money today!

Wayer2/
LDA-robocasa

Name: LDA-robocasa
Availability: InStock

vision-language-action

robocasa

lda

apache-2.0

Views

No views yet

LDA Robocasa Model

This repository provides the LDA Robocasa checkpoint and the auxiliary files required for inference with the LDA codebase.

GitHub: https://github.com/jiangranlv/LDA-1B
Project Page: https://pku-epic.github.io/LDA/

Files

The Hugging Face model repository contains:

text

LDA-robocasa.pt
config.yaml
dataset_statistics.json

LDA-robocasa.pt: PyTorch checkpoint weights.
config.yaml: model configuration used to rebuild the LDA framework.
dataset_statistics.json: dataset normalization statistics used to un-normalize predicted actions during inference.

Required Local Directory Layout

The current LDA loader expects the .pt checkpoint to be placed inside a subdirectory, usually named checkpoints, while config.yaml and dataset_statistics.json must stay in the parent run directory.

After downloading the files, organize them locally as:

text

LDA-robocasa/
|-- config.yaml
|-- dataset_statistics.json
`-- checkpoints/
    `-- LDA-robocasa.pt

The checkpoint path passed to LDA should be:

LDA-robocasa/checkpoints/LDA-robocasa.pt

Why This Layout Is Needed

baseframework.from_pretrained() loads the checkpoint path and infers the run directory from it:

python

checkpoint_pt = Path(pretrained_checkpoint)
run_dir = checkpoint_pt.parents[1]

For example, if the checkpoint path is:

LDA-robocasa/checkpoints/LDA-robocasa.pt

then the inferred run directory is:

LDA-robocasa

The loader then expects to find:

text

LDA-robocasa/config.yaml
LDA-robocasa/dataset_statistics.json

If LDA-robocasa.pt is placed directly next to config.yaml and dataset_statistics.json, the loader will infer the wrong parent directory and fail to find the required files.

Download And Prepare

You can download the model repository with huggingface_hub:

python

from pathlib import Path
import shutil

from huggingface_hub import snapshot_download

repo_dir = Path(snapshot_download(repo_id="YOUR_ORG_OR_USERNAME/LDA-robocasa"))

ckpt_dir = repo_dir / "checkpoints"
ckpt_dir.mkdir(exist_ok=True)

src_ckpt = repo_dir / "LDA-robocasa.pt"
dst_ckpt = ckpt_dir / "LDA-robocasa.pt"

if src_ckpt.exists() and not dst_ckpt.exists():
    shutil.move(str(src_ckpt), str(dst_ckpt))

print("Checkpoint path:", dst_ckpt)

Replace YOUR_ORG_OR_USERNAME/LDA-robocasa with the actual Hugging Face repository ID.

Load The Model

python

from lda.model.framework.base_framework import baseframework

ckpt_path = "LDA-robocasa/checkpoints/LDA-robocasa.pt"

model = baseframework.from_pretrained(ckpt_path)
model = model.to("cuda").eval()

Start The Policy Server

From the LDA repository root, run:

bash

python deployment/model_server/server_policy.py \
  --ckpt_path LDA-robocasa/checkpoints/LDA-robocasa.pt \
  --port 10093 \
  --use_bf16

Run RoboCasa Evaluation

In a separate terminal with the RoboCasa environment activated, run:

bash

export PYTHONPATH=$(pwd):${PYTHONPATH}

python examples/Robocasa_tabletop/eval_files/simulation_env.py \
  --args.env_name ${env_name} \
  --args.port 10093 \
  --args.n_episodes 50 \
  --args.n_envs 1 \
  --args.max_episode_steps 720 \
  --args.n_action_steps 12 \
  --args.video_out_path ${video_out_path} \
  --args.pretrained_path LDA-robocasa/checkpoints/LDA-robocasa.pt

You can also use the batch evaluation script:

bash examples/Robocasa_tabletop/eval_files/batch_eval_args.sh

Make sure the checkpoint path used by the script points to:

LDA-robocasa/checkpoints/LDA-robocasa.pt

Required Files Checklist

Before running inference, confirm that the following files exist:

text

LDA-robocasa/config.yaml
LDA-robocasa/dataset_statistics.json
LDA-robocasa/checkpoints/LDA-robocasa.pt

The checkpoint file must:

exist locally
use the .pt suffix
be placed one directory below the run directory

The config and statistics files must:

be named exactly config.yaml and dataset_statistics.json
be located in the inferred run directory
correspond to the same training run as the checkpoint

Troubleshooting

Missing `config.yaml`

If you see an error similar to:

Missing `config.yaml`

check that your local directory is organized as:

text

LDA-robocasa/
|-- config.yaml
|-- dataset_statistics.json
`-- checkpoints/
    `-- LDA-robocasa.pt

and that you pass:

LDA-robocasa/checkpoints/LDA-robocasa.pt

instead of:

LDA-robocasa/LDA-robocasa.pt

Missing `dataset_statistics.json`

If you see an error similar to:

Missing `dataset_statistics.json`

make sure dataset_statistics.json is in the same directory as config.yaml, not inside the checkpoints directory.

Invalid Checkpoint Suffix

The loader asserts that the checkpoint suffix is .pt. Make sure the checkpoint file is named:

LDA-robocasa.pt

Notes

dataset_statistics.json is required for action un-normalization. Removing or replacing it can cause predicted actions to be scaled incorrectly.

config.yaml is required because the LDA framework is rebuilt from the saved configuration before loading the checkpoint weights.

Wayer2/LDA-robocasa

LDA Robocasa Model

Files

Required Local Directory Layout

Why This Layout Is Needed

Download And Prepare

Load The Model

Start The Policy Server

Run RoboCasa Evaluation

Required Files Checklist

Troubleshooting

Missing config.yaml

Missing dataset_statistics.json

Invalid Checkpoint Suffix

Notes

Wayer2/
LDA-robocasa

Missing `config.yaml`

Missing `dataset_statistics.json`