---
license: mit
language:
- en
tags:
- cvrp
- vehicle-routing
- drone
- attention-model
- pomo
- onnx
- optimization
library_name: onnx
---

# Drone CVRP Neural Solver (Attention Model + POMO)

ONNX-exported neural network for solving Capacitated Vehicle Routing Problems (CVRP) in drone delivery contexts.

## Architecture

- **Encoder**: 6-layer Transformer with 8-head multi-head attention (128-dim embeddings, 512-dim FFN)
- **Decoder**: Autoregressive pointer-network with glimpse mechanism (Kool et al. 2019)
- **Input features**: 4D per node — x, y, z (altitude), demand
- **Training**: REINFORCE with greedy rollout baseline, 1000 epochs on random CVRP-50 instances

The model supports **dynamic batch and problem sizes** up to 200 nodes.

## ONNX Interface

| Port | Name | Shape | Type |
|------|------|-------|------|
| Input | `locs` | `[batch, nodes, 3]` | float32 |
| Input | `demand` | `[batch, nodes, 1]` | float32 |
| Input | `capacity` | `[batch, 1]` | float32 |
| Output | `actions` | `[batch, max_steps]` | int64 |
| Output | `log_p` | `[batch, max_steps]` | float32 |

The wrapper automatically prepends a depot at `(0.5, 0.5, 0.0)` with zero demand. Set `capacity` high (e.g. 999) for unconstrained TSP optimization, or use the trained value (40) for standard CVRP-50 benchmark routing.

## Training Recipe

| Parameter | Value |
|-----------|-------|
| Problem size | 50 customers |
| Batch size | 512–1024 |
| Epochs | 500–1000 |
| Learning rate | 1e-4 (cosine annealing) |
| Warmup | 10 epochs |
| Demand range | {1..9} |
| Vehicle capacity | 40 |
| Entropy bonus | 0.01 |

## Usage

```python
import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("cvrp50_model.onnx")

locs = np.random.rand(1, 10, 3).astype(np.float32)
demand = np.ones((1, 10, 1), dtype=np.float32)
capacity = np.array([[999.0]], dtype=np.float32)

actions, log_p = session.run(None, {
    "locs": locs,
    "demand": demand,
    "capacity": capacity,
})

print(actions[0, :20])
```

## Demo

Try it live in the [Drone VRP Space](https://aerialblancaservices-drone-vrp.hf.space) — side-by-side comparison with a nearest-neighbor heuristic, physics-based energy modeling, and interactive route visualization.

## References

- Kool, van Hoof, Welling (2018). *Attention, Learn to Solve Routing Problems!* ([arXiv:1803.08475](https://arxiv.org/abs/1803.08475))
- Kwon et al. (2020). *POMO: Policy Optimization with Multiple Optima for Reinforcement Learning* ([arXiv:2010.16011](https://arxiv.org/abs/2010.16011))
- Dorling et al. (2017). *Vehicle Routing Problems for Drone Delivery* ([arXiv:1609.02906](https://arxiv.org/abs/1609.02906))