Crazyflie Case Study

Crazyflie Case Study by HHU STUPS Group

Python JSON RPC Server with Crazyflie

This repository contains the source code for a Python JsonRPC server implementation. Furthermore, this repository provides some functionality of the crazyflie python library via ZMQ and the B Machines that can interface with it.

The code for the JSON RPC server is implemented in the files zmq_rpc.py and cf_rpc.py.

Getting started

  1. Install ProB2-UI in its latest nightly version.
  2. Start Crazyflie server
python cf_rpc.py
  1. Load the Drone.prob2project and load DroneMainController.mch

Step 1: Open Project

For animation:

  1. Perform SETUP_CONSTANTS, INITIALISATION.
  2. Perform other flying actions MAIN_TAKEOFF, MAIN_FORWARD, MAIN_OBSERVE etc.

Step 2a: Animation via Operations View

For autonomous execution as Hardware-in-the-loop Execution or Simulation:

  1. Open SimB in ProB2-UI

Step 2b: Open SimB

  1. Load RL Agent `DroneEnv.py

Step 3b: Load RL Agent

  1. Start the RL agent as a SimB simulation by clicking on the Start button

Step 4b: Start Simulation/Hardware-in-the-loop Execution

Real World Scenarios

HTML Scenarios with Videos

Scenario 3

Scenario 22

More Scenarios

Scenario 1

Scenario 2

Scenario 4

Scenario 5

Scenario 6

Scenario 7

Scenario 8

Scenario 9

Scenario 10

Scenario 11

Scenario 12

Scenario 13

Scenario 14

Scenario 17

Scenario 18

Scenario 19

Scenario 20

Scenario 21

Scenario 22

Example

State before action to fly backward:

State before action to fly backward

State after action to fly backward:

State after action to fly backward

Interaction with Crazyflie Server

One can interact with the Crazyflie server through the DroneCommunicator.mch B model, which can be loaded in the formal method tool ProB2-UI (https://github.com/hhu-stups/prob2_ui). The available commands to control the drones are as follows:

Command Description
Init Initialize communication with Crazyflie server
Destroy Destroys connection to socket
open_link Opens connection to a drone
close_link Closes connection to a drone
register_sensors Registers sensors of drone
Drone_Takeoff Sends command to drone for taking off
Drone_Land Sends command to drone to land
Drone_Left(dist) Sends command to drone to fly left
Drone_Right(dist) Sends command to drone to fly right
Drone_Up(dist) Sends command to drone to fly upward
Drone_Downward(dist) Sends command to drone to fly downward
Drone_Forward(dist) Sends command to drone to fly forward
Drone_Backward(dist) Sends command to drone to fly backward
Drone_GetLeftDistance Read distance sensor to the left
Drone_GetRightDistance Read distance sensor to the right
Drone_GetUpDistance Read distance sensor to the up
Drone_GetDownDistance Read distance sensor to the down
Drone_GetForwardDistance Read distance sensor to the forward
Drone_GetBackwardDistance Read distance sensor to the backward
Drone_GetX Read x position of drone
Drone_GetY Read y position of drone
Drone_GetZ Read z position of drone

Evaluation

In the following, we show the results of the RL agent in the simulation and compare them to the actual performance in the real-world, demonstrating the simulation-to-reality gap.

Performance of RL Agent in Simulation

Metric Unshielded Agent Shielded Agent
Safety-Critical Metrics
Mission failure [%] 2.9% 0.0%
Mission successful [%] 24.9% 24.1%
Mission fail-safe (with base return) [%] 44.1% 47.9%
Mission fail-safe (without base return) [%] 28.1% 28.0%
Performance Metrics
Actions performed 129.31 ± 56.81 130.65 ± 61.06
Mission coverage [%] 93.72% ± 9.55% 92.92% ± 11.41%
Total reward 1004.00 ± 715.28 1010.40 ± 674.17

This evaluation results from Monte Carlo simulation. The corresponding traces are available in Shielded Agent and Unshielded Agent, respectively. The Monte Carlo simulation as well as the traces are accessible through the project Drone.prob2project, which can be opened in ProB2-UI

Performance of RL Agent in Real World

This evaluation results from running the RL agent in the real world. The corresponding traces are available in Real Traces (some with videos). The traces are accessible through the HTML export and the project Drone.prob2project, which can be opened in ProB2-UI. One can re-run those traces through the B machine configured for replay DroneMainController_replay

Metric Shielded Agent
Mission Performance
Mission coverage [%] 54.60% ± 18.93%
Exploration Statistics
Position changes / Position updates 49 / 200 (24.50%)
Re-explored fields / Observed fields 428 / 665
Inconsistent detections / Re-explored fields 28 / 428 (6.5%)