Crazyflie autonomous mission

Final project for the EPFL course "Aerial Robotics"

Slides
Code

At a glance

This is the final project for the EPFL course “Aerial Robotics” (Prof. D. Floreano), completed in Spring 2022 in collaboration with Nay Abi Akl, Mariam Hassan, and Yehya El Hassan.

The project, “CrazyPractical,” consisted of implementing the onboard software for the Crazyflie quadrotor to autonomously:

  1. Take off from a starting pad
  2. Navigate through an arena while avoiding obstacles
  3. Find and localize a landing pad, then land safely
  4. Take off again from the landing pad
  5. Return to the initial pad and land, all within a strict time limit

This required a full autonomy stack combining planning, reactive avoidance, and robust localization under changing arena configurations.

System in a nutshell

The pipeline follows a pragmatic “search → avoid → confirm → act” structure, split into two mission phases.

Phase 1: Start pad → Landing pad

  1. Take-off from the take-off pad
  2. Sweep to find a free corridor / direction of motion
  3. Avoid obstacles during navigation (reactive layer)
  4. Sweep to detect the landing pad
  5. Landing-pad localization manoeuvre, then controlled landing

Phase 2: Landing pad → Start pad (return)

  1. Take-off from the landing pad
  2. Choose a return-planning strategy:
    • A* (global planning), or
    • Follow a forward path heuristic, or
    • Re-discretize a new path based on current conditions
  3. Take-off pad localization manoeuvre, then landing
Left: Phase 1 of the pipeline: Start → Sweep → Avoid → Locate → Land. Right: Phase 2 of the pipeline: Take-off → Plan return (A* or heuristic) → Locate start → Land.

What we implemented

Autonomous navigation with obstacle avoidance

A key requirement was safe navigation in cluttered environments. The solution combines:

  • A high-level navigation strategy (sweep-based exploration to progress and to search for pads)
  • A reactive obstacle avoidance layer to handle nearby obstacles and last-moment corrections

This hybrid approach kept the drone moving toward objectives while remaining safe when the environment changed.

Pad search and robust localization manoeuvres

Landing is only possible if the pad pose is estimated reliably. We implemented dedicated manoeuvres to:

  • Confirm pad detection
  • Reduce ambiguity through controlled motion (rather than trusting a single noisy observation)
  • Align for stable descent and touchdown

A similar manoeuvre is used again during the return phase to localize the original take-off pad before landing.

Return-path planning options (A* / heuristics / re-discretization)

For the second half of the mission, we included alternative strategies to plan the way back:

  • A* for a more globally consistent path when conditions allow it
  • Simpler forward/heuristic approaches when speed or reliability is preferable
  • Re-discretization to recompute a feasible path from the current state

This makes the system more flexible across different obstacle/pad configurations.

Results (from our trials)

From the test summary we performed:

  • Obstacle avoidance: 10/10 trials
  • Landing on landing pad: 9/10 trials
  • Landing back on take-off pad: 8/10 trials
  • Time limit respected: 8/10 trials
  • Average completion time: 1:51 minutes

These numbers reflect end-to-end performance, including both landings and the full round trip.

Demo video

This demo video shows the full mission: take-off → obstacle-aware navigation → landing-pad acquisition & landing → re-takeoff → return planning → final landing, completed in under two minutes on successful runs.

Acknowledgements

Course: EPFL — Aerial Robotics (Prof. D. Floreano)
Team: Nay Abi Akl, Mariam Hassan, Yehya El Hassan, Luca Zunino
Source of the cover image: Matin, drone detection Dataset, Roboflow | CC BY 4.0