Tools for RU-VLA

Avalilable Datasets

Combined Le-Robot Datasets

We present a consolidated community dataset for Le-Robot embodiment that integrates 598 open-source community datasets into a single unified corpus. The dataset comprises 22,709 episodes and approximately 9.4 million frames spanning 563 distinct tasks. Two versions of the dataset are available: one with task annotations in English and another with task annotations translated into Russian.

Libero-RU Dataset

We also open-source a translated to Russian language version of Libero dataset.

Setup Training On Le-Robot Datasets

Clone this repository

git clone https://github.com/dunnolab/vla-arena-oss.git
cd vla-arena-oss
export VLA_ARENA_OSS_PWD=$(pwd)
cd ../

Clone Open-Pi from the official repository

git clone --recurse-submodules https://github.com/Physical-Intelligence/openpi.git
cd openpi
git reset --hard 95aadc6b16d170e4b13ab2e0ac64fbf2d1bb8e31
GIT_LFS_SKIP_SMUDGE=1 git submodule update --init --recursive

Patch policy and config files

cp $VLA_ARENA_OSS_PWD/openpi-patches/policies/vla_arena_policy.py src/openpi/policies/
cp $VLA_ARENA_OSS_PWD/openpi-patches/training/config.py src/openpi/training

Install patched repository

uv venv .venv && GIT_LFS_SKIP_SMUDGE=1 uv pip install -e . --python .venv/bin/python

uv sync

cp -r ./src/openpi/models_pytorch/transformers_replace/* .venv/lib/python3.11/site-packages/transformers/

Using our models

Choose <model-name> depending on your purposes:

• pi05-libero-ru-en - pi05 checkpoint from Physical-Intelligence finetuned on both Libero-EN and Libero-RU Datasets.
• pi05-pnp-ru - pi05 checkpoint from Physical-Intelligence finetuned on the pick-and-place dataset for SO101 Robot with Russian task annotations.

HF repository contains both model parameters and global normalization statistics which were used to train our models.

Load model from HF repo

# Make sure the hf CLI is installed
curl -LsSf https://hf.co/cli/install.sh | bash

# Download model from 
huggingface-cli download dunnolab/<model-name> --local-dir <your-local-dir> --local-dir-use-symlinks False
# You will get parameters (either `params/` folder or `model.safetensors` file)
# and `assets/` folder with normalization statistics 
# that were used during training and inference

Run finetuning

Before you start a traininig, it is important to copy normalization statistics (supposing you run commands from original openpi directory)

cp <your-local-dir>/assets/norm_stats.json assets/<config_name>/<repo_id>/
# correct path example: assets/pi05_libero_ru/dunnolab/libero_ru/norm_stats.json

We provide the following configs for training:

• pi05_so_combined_eng - to train pi05 on dunnolab/so-combined-eng
• pi05_so_combined_ru - to train pi05 on dunnolab/so-combined-ru
• pi05_libero_ru - to train pi05 on dunnolab/libero_ru
• pi0_libero_ru - to train pi0 on dunnolab/libero_ru
• pi0_so_combined_ru - to train pi0 on dunnolab/so-combined-ru
• pi0_so_combined_eng - to train pi0 on dunnolab/so-combined-eng

uv run torchrun \
  --standalone \
  --nnodes=1 \
  --nproc_per_node=8 \
  scripts/train_pytorch.py pi05_libero_ru \
  --project-name openpi-ru \
  --exp-name pi05-libero-ru-experiment \
  --num-train-steps 1000001 \
  --save-interval 5000 \
  --pytorch-weight-path /path/to/your/folder/with/pytorch/model


# if you want to start training in jax from a specified checkpoint -
# you should change a path in `weight_loader` of `TrainConfig`
export XLA_PYTHON_CLIENT_MEM_FRACTION=0.9
uv run scripts/train.py pi0_so_combined_eng \
  --exp-name=my_experiment \
  --overwrite

Run evaluation on LIBERO RU/ENG

We provide the following LIBERO-RU tasks:

• ru_libero_spatial - 10 tasks of libero_saptial (Russian)
• ru_libero_object - 10 tasks of libero_object (Russian)
• ru_libero_goal - 10 tasks libero_goal (Russian)
• ru_libero_10 - 10 tasks of libero_10 (Russian)

If you're using LIBERO as a third-party dependency (e.g., in the OpenPI project), you'll need to copy the Russian task files into your third_party/libero directory. Before that please refer to the installation instruction of the original openpi repository

The bddl files define the task specifications and init files contain the initial environment states in Russian. python files configure benchmark suite for compatibility of original libero repo with Russian tasks. Copy those files from your appropriate directory to your third-party LIBERO installation:

cp -r $VLA_ARENA_OSS_PWD/libero-ru/bddl_files/ru_libero_* third_party/libero/libero/libero/bddl_files/

cp -r $VLA_ARENA_OSS_PWD/libero-ru/init_files/ru_libero_* third_party/libero/libero/libero/init_files/

cp $VLA_ARENA_OSS_PWD/libero-ru/benchmark/libero_suite_task_map.py third_party/libero/libero/libero/benchmark/libero_suite_task_map.py

cp $VLA_ARENA_OSS_PWD/libero-ru/benchmark/__init__.py third_party/libero/libero/libero/benchmark/__init__.py

cp $VLA_ARENA_OSS_PWD/libero-ru/examples/main.py examples/libero/main.py

Then you can follow the original setup of openpi LIBERO installation:

...
# steps from the original openpi libero installation
...

uv pip install -e packages/openpi-client
uv pip install -e third_party/libero
export PYTHONPATH=$PYTHONPATH:$PWD/third_party/libero

Then, as an example, you can use such a script for LIBERO-RU/EN evaluation

# ru_libero_spatial ru_libero_object ru_libero_goal ru_libero_10
python examples/libero/main.py \
    --args.num_trials_per_task 50 \
    --args.task-suite-name ru_libero_spatial \
    --args.port 8000