Learning Parameterized Skills from Demonstrations (NeurIPS 2025)

This repository contains the official implementation of DEPS, an end-to-end algorithm for discovering parameterized skills from expert demonstrations. DEPS jointly learns discrete skills, their continuous parameters, and corresponding low-level action policies through a hierarchical architecture, combining the structure of discrete skills with the flexibility of continuous conditioning.

paper: https://openreview.net/pdf?id=FiZ7gadynD

website: https://sites.google.com/view/parameterized-skills

Installation

Prerequisites

• Python 3.8+
• CUDA 11.3+ (for GPU support)

Setup Steps

1. Clone the repository

   git clone https://github.com/guptbot/DEPS.git
   cd DEPS

2. Create and activate virtual environment

   python3.9 -m venv env
   source env/bin/activate  # On Windows: env\Scripts\activate

3. Install PyTorch with CUDA 11.3

   pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0+cu113 --extra-index-url https://download.pytorch.org/whl/cu113

4. Install all dependencies

   pip install -r requirements.txt

   This will automatically install LIBERO from GitHub along with all other required packages.

5. Configure LIBERO path

   Update the libero_root_path in your command or in config.py:

   --libero_root_path /path/to/libero/directory/

6. Download LIBERO datasets

   Follow instructions at https://github.com/Lifelong-Robot-Learning/LIBERO to download the required datasets (e.g., LIBERO-100). Make sure the downloaded datasets are stored in libero/libero/datasets

Usage

NOTE: the main branch of this repository contains the implementation of DEPS for use with LIBERO. To use DEPS with MetaWorld, checkout branch metaworld.

1. Pretraining:

   To pretrain DEPS:

   python main.py --train 1 --name run_name --batch_size 3 --seed 95 --markov_subpolicy 1 --train_dataset libero_90 --test_dataset libero_90  --algo deps --apply_aug 1 --number_policies 20 --hidden_size 1024 --hidden_size_primitive 1024 --hidden_size_compression 1024 --epochs 21 --z_dimensions 4 --kl_loss_weight 0.5
   

2. Finetuning:

   To finetune the checkpoint from step (1), run:

   python main.py --finetune 1 --name run_name --batch_size 2 --seed 95 --markov_subpolicy 1 --train_dataset libero_90 --test_dataset libero_90 --algo deps --apply_aug 0 --number_policies 20 --finetune_subpolicy 1 --finetune_resnet 1 --kl_loss_weight 0 --z_loss_weight 0 --hidden_size 1024 --hidden_size_primitive 1024 --hidden_size_compression 1024 --epochs 20 --z_dimensions 4 --model /path/to/saved/checkpoint --use_corrected_action 0 --downstream_rollout_frequ 2
   

More details on the specific commands to recreate the various presented runs in the paper can be found in run_commands.md.

In our experiments, we perform pretraining on 2 NVIDIA GeForce RTX 3090 GPUs and finetuning on a single NVIDIA GeForce RTX 3090 GPU. For MetaWorld, we use a single GPU for both pretraining and finetuning.

Provided LfD Algorithms

• deps: DEPS with state compression - learns 1D compressed robot state representation
• deps_uncompressed: DEPS without compression - uses full robot state
• multitask: Baseline that trains a single policy for all tasks

Directory Structure

├── main.py                    # Entry point
├── config.py                  # Argument parser
├── train.py                   # Base training class (PolicyManagerBase)
├── deps.py                    # DEPS with compression
├── deps_uncompressed.py       # DEPS without compression
├── multitask.py               # Multitask baseline
├── networks/
│   ├── networks.py            # Core neural architectures
│   ├── gmm_head.py            # GMM action head
│   └── resnet.py              # ResNet encoder with FiLM
├── env/
│   ├── libero_dataloader.py  # Dataset loading
│   ├── libero_rollout.py     # Environment wrapper
│   └── libero_spatial_encode.py  # Observation encoding
└── utils/
    ├── utils.py               # Helper functions
    ├── logger.py              # Logging interface (WandB)
    └── visualization_tools.py # Visualization utilities

Citation

If you find this repository useful, please cite our work:

@inproceedings{vedant2025deps,
  title     = {Learning parameterized skills from demonstrations},
  author    = {Vedant Gupta and Haotian Fu and Calvin Luo and Yiding Jiang and George Konidaris},
  booktitle = {NeurIPS},
  year      = {2025},
}