Running Flocking Algorithm with 20 Crazyflies in Crazyswarm2

[jialongzeng]

Thank you for your excellent work on Crazyswarm2! I am currently attempting to run a flocking algorithm with 20 CFs. However, the best I have managed to achieve so far is with 12 CFs. Below are the details of my setup and the challenges I am facing.

---

Experiment Configuration:

1. Hardware Setup

• Crazyflies: 12 Crazyflie 2.1 units with the latest firmware.
• Crazyradio PAs: 4 units with firmware version 99.55.
  • Due to limited USB ports on my laptop:
    • 3 PAs are connected to a single USB 3.0 docking station.
    • 1 PA is connected to another USB docking station.
• Positioning System: NOKOV motion capture system streaming at 100Hz (adjustable).

2. Connection Setup

• Each PA connects to 3 CFs (grouped) on the same channel.
• Unique URI addresses are assigned to each CF.

3. Logging and Control

• Firmware Logging: Only odom is enabled at 10Hz.
• Control System:
  • All CFs are controlled by an individual ROS 2 Python node.
  • Nodes subscribe to all the odometry data (/cfxx/odom) and publish all the full-state control commands (/cfxx/cmd_full_state).
• Backend: Using cflib, since velocity data is required by the control algorithm and can be easily access through odometry.
• Environment: Ubuntu 22.04 and ROS2 running in Docker with privileged access for peripherals.

---

Observations:

1. Connection Quality:
   • The quality seems to depend significantly on the direction and distance between the PAs and CFs.
2. LED Behavior:
   • With fewer CFs connected, the LEDs on both the CFs and PAs glow strong green.
   • As the number of CFs and PAs increases:
     • CF LEDs shift from orange (possibly red + green) to fast-flashing week green.
     • PA LEDs flash with both green and red lights.

---

Questions:

1. Recommended Configurations:

   • What is your common setup for controlling a large number of CFs (e.g., backend, control paradigm, hardware settings like motion capture frequency, CF logging, and channel configurations)?
   • Are there any adjustments you would suggest to improve my experiment setup?

2. Factors Affecting Communication Quality:

   • What are the primary factors influencing communication quality?
   • Could the use of Docker or USB docking stations contribute to delays or degraded performance?

[whoenig]

For Crazyswarm2, I only tested up to 10 CFs with the Cpp backend. We typically use 100 Hz for the mocap, although 75 Hz is also working well especially if connecting many drones over one radio. For channels 80, 82, etc. work well in our case (and 2M datarate).

I don't think Docker or USB hubs are a particular problem. For the Cpp backend, we noticed that the chipset is important (Intel works better than AMD).

The cpp backend can measure latency (and so can the cflib now, but I don't think this was hooked up to the ROS backend yet, @knmcguire ?), and this measure might help you identify problems as you scale up.

[jialongzeng]

Thank you for your helpful response! I still have a few follow-up questions:

1. How many CFs are connected to a single radio in your setup? From other discussions, I've learned that 3-4 CFs per radio is generally recommended for the cflib backend.
2. Do you have any insights into factors that might limit scalability?
3. Finally, would using cpp instead of cflib as backend improve performance when connecting many CFs?
   Thanks in advance!

[whoenig]

1. I had 10 on one radio, but without any log-topics (apart from status). The more log topics you have the more radios you need. 4 for cflib should work well, though.
2. There are known issues with the cpp backend and lost packets. For the cflib backend, it should scale well as long as you use enough radios.
3. Improving performance in terms connecting for sure. The cflib also uses broadcasts, which significantly improves the performance for motion capture (only one packet for all drones need to be send, rather than 4; effectively reducing latency and increasing bandwidth). The downside is this lost packet issue mentioned in the second point. It seems to be fine on computers with Intel chipsets, but is very noticeable for machines with AMD chipset.

[knmcguire]

Hi just to append to the support, the crazyflie python library has now the link statistics merged as of end of last year: bitcraze/crazyflie-lib-python#492.

@ataffanel, I'm not sure if this will be the final state or if it is still quite in the breaking change realm. I guess it might be good to wait a bit until implementing this, or what do you think?

[knmcguire]

I've made an issue for it: #622