The environment used can be replicated as: conda env create -f beampy_nodep.yml

Data Generation can be done using BeamNG, by running
./GenerateData.sh

Training ZLIK: python3 TrainingPipeline/zlik_train.py

Based on the project structure and training scripts found in your files, here is a draft for your new project's README.md following the flow of your previous project:

📄 Checkout our paper: [https://arxiv.org/abs/2602.12385]

🎥 Checkout our Video: [(https://www.youtube.com/watch?v=QbtaDONTQdQ&)]

ZLIK enables adaptive kinodynamic modeling for robots with structural damage by leveraging natural language descriptions. This allows for zero-shot adaptation to various damage scenarios, ensuring stable motion planning even under unforeseen structural failures.

Traditional kinodynamic models often fail when a robot undergoes structural changes or damage. ZLIK (Zero-Shot Language-Informed Kinodynamics) addresses this by incorporating high-level natural language descriptions of the robot's state into the dynamics model.

Key features include:

Zero-Shot Adaptation: Generalizes to novel damage scenarios without retraining on specific failure cases.

Language Integration: Uses a language-informed approach to bias the dynamics model based on text embeddings (e.g., "front right tyre punctured").

Transformer-Based Architecture: Employs an encoder-decoder transformer to predict future robot poses based on historical states, future actions, and damage context.

The environment can be replicated using Conda:

conda env create -f beampy_nodep.yml
conda activate ros2_beampy

Data is generated using the BeamNG simulator, covering various damage scenarios like tyre punctures, axle breaks, and suspension failures.

Run the data generation pipeline:

./GenerateData.sh

This script handles:

1. Running random walks with/without damage.

2. Encoding damage embeddings via embedding-gemma.

3. Extracting trajectories for training.

ZLIK involves a multi-stage training process including pre-training and specialized dynamics training.

python3 TrainingPipeline/pre_train.py --conf config/pre_train_config.yaml

You can train specific dynamics models depending on the robot state:

Clean Dynamics:

python3 TrainingPipeline/clean_dynamics_train.py --conf config/damaged_model_config.yaml

ZLIK (Language-Informed):

python3 TrainingPipeline/zlik_train.py --conf TrainingPipeline/conf/damaged_model_config.yaml --encoder_conf config/pre_train_config.yaml

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

@misc{pokhrel2026zeroshot,
      title={Zero-Shot Adaptation to Robot Structural Damage via Natural Language-Informed Kinodynamics Modeling}, 
      author={Anuj Pokhrel and Aniket Datar and Mohammad Nazeri and Francesco Cancelliere and Xuesu Xiao},
      year={2026},
      eprint={2602.12385},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2602.12385}, 
}