---
license: mit
---
# Geometric Reasoning Networks (GRN)

![GRN model architecture](https://github.com/Smail8/geometric_reasoning_networks/blob/main/assets/model.png?raw=true)

## Model Details

- **Model name:** Geometric Reasoning Networks (GRN)
- **Model type:** Graph Neural Network for robot manipulation feasibility prediction
- **Framework:** PyTorch, PyTorch Geometric
- **Associated paper:**  
  *Learning Geometric Reasoning Networks for Robot Task and Motion Planning*, ICLR 2025
- **Authors:** Smail Ait Bouhsain, Rachid Alami, Thierry Siméon
- **License:** See repository LICENSE
- **Paper:** https://openreview.net/pdf?id=ajxAJ8GUX4
- **Repository:** https://github.com/smail8/geometric_reasoning_networks

GRN is a learned geometric reasoning model designed to augment **robot Task and Motion Planning (TAMP)** by predicting the feasibility of manipulation actions in cluttered 3D environments.

---

## Model Description

The Geometric Reasoning Network (GRN) operates on **graph-structured representations of manipulation scenes**, encoding objects, candidate grasps, actions, and their geometric relations. It predicts action and grasp feasibility by reasoning jointly over learned geometric constraints.

GRN integrates predictions from auxiliary learned modules:
- **Inverse Kinematics (IK) feasibility**
- **Grasp Obstruction (GO)**
- **Action and Grasp Feasibility (AGF)**

---

## Intended Use

### Primary Use Cases
- Augmenting robot Task and Motion Planning (TAMP)
- Learned feasibility prediction for manipulation actions
- Research in robotic manipulation and geometric reasoning
- Benchmarking learned planning heuristics

### Out-of-Scope Use
- Low-level control or trajectory optimization
- Safety-critical deployment without validation
- Non-robotic domains

---

## Model Architecture

- Graph Attention Network with message passing
- Nodes represent objects
- Edges encode geometric and relational constraints
- Supervised training on large-scale simulated datasets

---

## Training Details

- **Training data:** GRN simulated manipulation datasets
- **Loss:** Binary cross-entropy + Mean-squared error
- **Optimizer:** Adam
- **Hyperparameters:**
  | Module    | Batch size | Learning Rate | Epochs |
  | --------- | ---------- | ------------- | ------ |
  | IK        | 8192       | 1e-3          | 100    |
  | GO        | 8192       | 1e-3          | 100    |
  | AGF       | 2048       | 1e-4          | 100    |
  | GRN       | 2048       | 1e-4          | 100    |
 
- **Hardware:**  NVIDIA RTX A5000 GPU
- **Training Time:** 15 hours

---

## Evaluation

Evaluation is performed on both in-distribution and out-of-distribution datasets with increasing scene complexity.

Metrics include:
- Feasibility prediction accuracy
- Downstream task and motion planning success rate
- Generalization to unseen clutter levels and object sizes

GRN outperforms MLP, CNN-based, and standard GNN baselines. Please refer to the paper for detailed results.

![GRN results visualization](https://github.com/Smail8/geometric_reasoning_networks/blob/main/assets/viz.png?raw=true)

---

## Limitations

- Trained exclusively in simulation
- Performance degrades for extreme clutter or unseen geometries
- Assumes accurate scene geometry and object poses

---

## Ethical Considerations

This model does not involve human data. Users are responsible for ensuring safe deployment when integrated into physical robotic systems.

---

## Citation

```bibtex
@inproceedings{ait2025learning,
  title={Learning Geometric Reasoning Networks for Robot Task and Motion Planning},
  author={Ait Bouhsain, Smail and Alami, Rachid and Simeon, Thierry},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}