PURT-UAS Biomed Drone Delivery

Overview

PURT-UAS Biomed Drone Delivery is a prototype overdose-response drone focused on perception → safe landing decisions → navigation.
My work centered on depth + vision sensing and algorithmic decision-making: selecting safe landing zones from point clouds and exploring planning methods like RRT* / A* for real-time pathfinding.

What I Built

1) RealSense Depth + Frame Capture (Perception Input)

I implemented a RealSense capture module to stream aligned depth + color, then derived sensor signals like nearest obstacle distance from depth data.

import pyrealsense2 as rs
import numpy as np
import cv2

class RealSenseCapture:
    def __init__(self, width=1280, height=720, fps=30):
        self.pipeline = rs.pipeline()
        self.config = rs.config()
        self.config.enable_stream(rs.stream.depth, width, height, rs.format.z16, fps)
        self.config.enable_stream(rs.stream.color, width, height, rs.format.bgr8, fps)
        self.align = rs.align(rs.stream.color)

    def start(self):
        profile = self.pipeline.start(self.config)
        depth_sensor = profile.get_device().first_depth_sensor()
        self.depth_scale = depth_sensor.get_depth_scale()
        depth_sensor.set_option(rs.option.visual_preset, 3)  # accuracy preset

    def get_frames(self):
        frames = self.pipeline.wait_for_frames()
        aligned = self.align.process(frames)
        depth_frame = aligned.get_depth_frame()
        color_frame = aligned.get_color_frame()
        depth = np.asanyarray(depth_frame.get_data()).copy()
        color = np.asanyarray(color_frame.get_data()).copy()
        gray = cv2.cvtColor(color, cv2.COLOR_BGR2GRAY)
        return gray, depth

    def get_min_distance_cm(self):
        frames = self.pipeline.wait_for_frames()
        depth_frame = frames.get_depth_frame()
        depth_img = np.asanyarray(depth_frame.get_data())
        return float(np.min(depth_img * self.depth_scale) * 100.0)

Tech Stack

Perception & Sensing

Intel RealSense (pyrealsense2) — aligned depth + RGB streaming
OpenCV — image processing + preprocessing
NumPy — point cloud construction and geometry

Geometry & Decision Systems

RANSAC Plane Fitting — robust landing surface detection
Point Cloud Processing — terrain modeling + height maps
Geometric Metrics — slope, flatness, usable area, obstacle proximity

Planning & Autonomy

ROS — robotics integration + autonomy stack
RRT* — sampling-based motion planning
A* — graph-based path planning baseline
Graph Algorithms — navigation abstractions

Visualization & Simulation

Matplotlib (2D + 3D) — point clouds, height maps, landing zones
Synthetic Terrain Generators — realistic terrain + stress testing

Model Architecture

Perception → Geometry → Decision Pipeline

Results

Landing Zone Detection

Robust plane detection under noisy and uneven terrain using RANSAC
Stable LAND / ABORT decisions based on:
- slope thresholds
- surface flatness
- inlier density
- usable landing area
- obstacle proximity

Terrain Analysis

Generated realistic terrain with:
- hills
- depressions
- ridges
- plateaus
Validated landing detection across varied surface geometries using synthetic and semi-realistic terrain models

Planning

Demonstrated RRT* behavior:
- tree expansion
- rewiring optimization
- cost reduction over iterations
- feasible path extraction in continuous space
Established a planning baseline for real-time autonomous navigation

Impact

Converts raw depth + vision data into actionable autonomy decisions
Bridges perception, geometry, and planning into a unified pipeline
Provides a foundation for safe autonomous biomedical deployment
Demonstrates real-world robotics reasoning: not just models, but systems

What I Learned

How to turn noisy perception into reliable geometric decisions
Why RANSAC + inlier structure matters more than raw model fitting
Practical tradeoffs between A* (structured maps) vs RRT* (continuous space)
The importance of visualization-first debugging in robotics systems
Designing autonomy as a pipeline, not a single model