README.md

Deployment

This directory contains tools for deploying trained models on physical drones and running real-world tests.

Overview

The deployment module provides:

1. Model conversion: Export PyTorch models to C for embedded systems
2. Validation: Verify model correctness after conversion
3. Ivy senders: Stream evader trajectories to Paparazzi autopilot for testing

Model Deployment

Extracting Models to C

File: extract_validate.py

This script converts trained PPO models to C headers for embedded deployment:

1. Extracts weights from PyTorch model state dict
2. Generates C header (ppo_weights.h) with weight arrays
3. Compiles shared library (libppo_controller.so or .dylib)
4. Validates by comparing PyTorch and C outputs

Usage:

python deployment/extract_validate.py \
    --model path/to/trained/model.zip \
    --csrc deployment/ppo_controller.c

Note, ensure that the input and output dimensions in the extract validate script match that of ppo_controller.c.

Configuration:

HEADER_NAME = "ppo_weights.h"
LIB_BASE = "libppo_controller"
TOLERANCE = 1e-5              # Validation tolerance
NUM_TESTS = 500               # Number of validation tests
OBS_RANGE = (-100.0, 100.0)   # Observation range for testing
INPUT_DIM = 18                # Observation dimension 
OUTPUT_DIM = 3                # Action dimension

Output:

• ppo_weights.h: C header with network weights
• libppo_controller.{so,dylib}: Compiled shared library

Validation: The script runs random test observations through both PyTorch and C implementations and checks outputs match within tolerance. Exit code 0 = success, 1 = failure.

C Controller Template

File: ppo_controller.c

Template C implementation of a PPO policy forward pass. You'll need to:

1. Copy this file to your build directory
2. Include the generated ppo_weights.h
3. Implement the forward pass using the weight arrays
4. Compile into a shared library

The template assumes:

• ReLU activations
• Linear layers
• Tanh output activation (for bounded actions)

Key Functions:

• ppo_forward(): Main forward pass function
• Input: observation array
• Output: action array

Using Ivy Senders

See ivy_senders/README.md for detailed documentation on streaming evader trajectories.