RoboMD: Uncovering Robot Vulnerabilities through Semantic Potential Fields

April 12, 2026 · View on GitHub

We introduce RoboMD a deep reinforcement learning-based framework designed to identify failure modes in robotic manipulation policies. By simulating diverse conditions and quantifying failure probabilities, RoboFail provides insights into model robustness and adaptability.

Installation

Prerequisites

Ensure you have the following dependencies installed:

Python 3.8+
CUDA (if using GPU)
Conda (recommended for managing environments)

Setting Up the Environment

Clone the repository:

git clone https://github.com/Robo-MD/Robo-MD-RSS.github.io.git
cd Robo-MD-RSS.github.io

Create a Conda environment:

conda create --name robomd python=3.8 -y
conda activate robomd

Installing Dependencies

1. Install robosuite

pip install robosuite

2. Install robomimic

git clone https://github.com/ARISE-Initiative/robomimic.git
cd robomimic
pip install -e .

3. Additional Dependencies

Install required Python packages:

pip install -r requirements.txt

Project Structure

├── configs/               # Configuration files for actions and training
├── env/                   # Environment implementations
├── scripts/               
├── utils/                 # Utility functions (e.g., loss computations)
├── train_continuous.py     # Training script for continuous latent actions
├── train_discrete.py       # Training script for discrete latent actions
├── train_embedding.py      # Training script for embedding learning
├── README.md               # Project documentation
├── requirements.txt        # Required dependencies

Usage

Training with Continuous Actions

To train an RL policy using a latent action space, run:

python train_continuous.py --name <run_name> --task <task_name> --agent <path_to_agent> --rl_timesteps 3000

Example:

python train_continuous.py --name latent_rl --task lift --agent models/bc_agent.pth --rl_timesteps 50000

Training Discrete Action Policies

For training RL with a discrete action space:

python train_discrete.py --name <run_name> --task <task_name> --agent <path_to_agent> --rl_timesteps 3000

Training Embeddings

To train and store known embeddings:

python train_embedding.py --path <dataset_path>

This script extracts embeddings from a dataset and stores them in an HDF5 file.