Repository: text2nav

Accepted to Robotics: Science and Systems (RSS) 2025 Workshop on Robot Planning in the Era of Foundation Models (FM4RoboPlan)

This repository contains the implementation for our research investigating whether frozen vision-language model embeddings can guide robot navigation without fine-tuning or specialized architectures. We present a minimalist framework that achieves 74% success rate in language-guided navigation using only pretrained SigLIP embeddings.

• 🎯 74% success rate using frozen VLM embeddings alone (vs 100% privileged expert)
• 🔍 3.2x longer paths compared to privileged expert, revealing efficiency limitations
• 📊 SigLIP outperforms CLIP and ViLT for navigation tasks (74% vs 62% vs 40%)
• ⚖️ Clear performance-complexity tradeoffs for resource-constrained applications
• 🧠 Strong semantic grounding but limitations in spatial reasoning and planning

Our approach consists of two phases:

1. Expert Demonstration Phase: Train a privileged policy with full state access using PPO
2. Behavioral Cloning Phase: Distill expert knowledge into a policy using only frozen VLM embeddings

The key insight is using frozen vision-language embeddings as drop-in representations without any fine-tuning, providing an empirical baseline for understanding foundation model capabilities in embodied tasks.

• Python 3.8+
• NVIDIA Isaac Sim/Isaac Lab
• PyTorch
• CUDA-compatible GPU

git clone https://github.com/oadamharoon/text2nav.git
cd text2nav

# Install dependencies
pip install torch torchvision
pip install transformers
pip install numpy matplotlib
pip install gymnasium

# For Isaac Lab simulation (follow official installation guide)
# https://isaac-sim.github.io/IsaacLab/

text2nav/
├── CITATION.cff           # Citation information
├── LICENSE                # MIT License
├── README.md              # This documentation
├── IsaacLab/              # Isaac Lab simulation environment setup
├── embeddings/            # Vision-language embedding generation
├── rl/                    # Reinforcement learning expert training
├── generate_embeddings.ipynb    # Generate VLM embeddings from demonstrations
├── revised_gen_embed.ipynb      # Revised embedding generation
├── train_offline.py             # Behavioral cloning training script
├── offlin_train.py              # Alternative offline training
├── bc_model.pt                  # Trained behavioral cloning model
├── td3_bc_model.pt            # TD3+BC baseline model
├── habitat_test.ipynb         # Testing and evaluation notebook
└── replay_buffer.py           # Data handling utilities

cd rl/
python train_expert.py --env isaac_sim --num_episodes 500

jupyter notebook generate_embeddings.ipynb

python train_offline.py --model siglip --embedding_dim 1152 --batch_size 32

• Environment: 3m × 3m arena in NVIDIA Isaac Sim
• Robot: NVIDIA JetBot with RGB camera (256×256)
• Task: Navigate to colored spheres based on language instructions
• Targets: 5 colored spheres (red, green, blue, yellow, pink)
• Success Criteria: Reach within 0.1m of correct target

1. Semantic Grounding: Pretrained VLMs excel at connecting language descriptions to visual observations
2. Spatial Limitations: Frozen embeddings struggle with long-horizon planning and spatial reasoning
3. Prompt Engineering: Including relative spatial cues significantly improves performance
4. Embedding Dimensionality: Higher-dimensional embeddings (SigLIP: 1152D) outperform lower-dimensional ones

• Hybrid architectures combining frozen embeddings with lightweight spatial memory
• Data-efficient adaptation techniques to bridge the efficiency gap
• Testing in more complex environments with obstacles and natural language variation
• Integration with world models for better spatial reasoning

@misc{subedi2025pretrainedvisionlanguageembeddingsguide,
      title={Can Pretrained Vision-Language Embeddings Alone Guide Robot Navigation?}, 
      author={Nitesh Subedi and Adam Haroon and Shreyan Ganguly and Samuel T. K. Tetteh and Prajwal Koirala and Cody Fleming and Soumik Sarkar},
      year={2025},
      eprint={2506.14507},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2506.14507}, 
}

This work is funded by NSF-USDA COALESCE grant #2021-67021-34418. Special thanks to the Iowa State University Mechanical Engineering Department for their support.

• Nitesh Subedi* (Iowa State University)
• Adam Haroon* (Iowa State University)
• Shreyan Ganguly (Iowa State University)
• Samuel T.K. Tetteh (Iowa State University)
• Prajwal Koirala (Iowa State University)
• Cody Fleming (Iowa State University)
• Soumik Sarkar (Iowa State University)

*Equal contribution

This project is licensed under the MIT License - see the LICENSE file for details.

• Paper (arXiv)
• RSS 2025 FM4RoboPlan Workshop
• Iowa State University

For questions or issues, please open a GitHub issue or contact the authors.