Open Source RoboticsRobots, software & learning

PULP-Dronet

Software

Deep-learning visual navigation engine for the Crazyflie 2.1 nano-drone: PyTorch CNNs, quantization and deployment flow for the AI-deck GAP8 processor, onboard flight application and a collected dataset; code Apache-2.0, v3 dataset CC BY-NC-SA 4.0. [1] [3] [4]

Installation & first result →Official websiteGetting started

For
Embedded-AI and nano-drone researchers able to train PyTorch models and flash Crazyflie and GAP8 firmware [4]
Hardware
Bitcraze Crazyflie 2.1 with AI-deck (GAP8, Himax camera) for v2/v3, PULP-Shield on Crazyflie 2.0 for v1 [1] [4]
Source rights
Named licenses are documented by part. Includes noncommercial terms. Review licenses
Unknowns
operating system not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
Apache-2.0, CC BY-NC-SA 4.0 (noncommercial) · Details by part

Requirements · Research details and dates

On this page

Overview

PULP-Dronet, from the PULP Platform group at ETH Zurich and the University of Bologna, is a set of convolutional neural networks with the code to train them, compress them and run them aboard a nano-drone. The README describes it as a visual navigation engine that lets a pocket-size quadrotor avoid collisions, including with moving obstacles, without a remote laptop. Its software builds on the DroNet network from the University of Zurich. This is a model and toolchain project rather than a drone design. The aircraft is a Bitcraze Crazyflie with a camera-and-processor expansion board. [1]

Is this for you?

Workflows
  • Dataset collection
  • Training
  • Testing
  • Quantization (NEMO)
  • Deployment (DORY)
  • Onboard autonomous flight

[4] [5]

Known limitations
  • Deployment pinned to crazyflie-firmware 2022.01 and gap-sdk 3.9.1; upstream warns new firmware versions are incompatible
  • New AI-decks may need specific NINA firmware
  • v3 dataset is noncommercial (CC BY-NC-SA 4.0) and downloaded separately
  • Submodules (NEMO, DORY, viewer, modified firmware) not included in the archive

Source-review boundaries; no training or flight performed.

[3] [4] [6] · Editorial assessment

What you need

Languages / runtimes
  • Python
  • C

PyTorch training scripts and C applications for STM32 and GAP8.

[4] [6] · Editorial assessment

Dependencies
  • Conda environment (conda_deps.yml) with PyTorch
  • NEMO quantization tool
  • DORY deployment tool (dory_dronet)
  • GreenWaves gap-sdk 3.9.1
  • crazyflie-firmware tag 2022.01
  • crazyflie-clients-python 2022.01 and crazyflie-lib-python 0.1.17.1

[4]

Compute for the workload
  • GPU for training (scripts take --gpu)
  • GAP8 SoC on the AI-deck for inference
  • JTAG cable for flashing GAP8

From documented commands and flashing steps.

[4] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [4]

  • Clone recursively to fetch NEMO, DORY and other submodules
  • Create the conda environment from conda_deps.yml
  • Download the v3 dataset from Zenodo for training or testing
  • Install gap-sdk 3.9.1 and crazyflie-firmware tag 2022.01
  • Flash the Crazyflie application and the GAP8 application, then start via the cfclient fly parameter

Editorial source-navigation sequence, not a reproduced or safety-qualified procedure.

[4] · Editorial assessment

Installation

Recursive git clone, conda environment, gap-sdk 3.9.1, Crazyflie firmware 2022.01; flash STM32 app with make cload and GAP8 app over JTAG.

[4]

Hardware, models & data

Supported hardware
  • Bitcraze Crazyflie 2.1 with AI-deck (GAP8, Himax camera) for v2/v3
  • PULP-Shield on Crazyflie 2.0 for v1

[1] [4]

Inputs and outputs

Grayscale Himax camera images in; collision prediction and steering (yaw-rate) outputs passed over UART to the Crazyflie flight application.

Synthesized from dataset labels, testing metrics (collision accuracy, yaw-rate RMSE) and application description.

[4] · Editorial assessment

File formats
  • PyTorch .pth weights
  • ONNX export from NEMO
  • JPEG images with CSV labels

[4]

Algorithms and models
  • PULP-Dronet v3 (ResBlock, depth multiplier 1.0)
  • Tiny-PULP-Dronet v3 (depthwise/pointwise, depth multiplier 0.125)
  • Earlier v1 and v2 DroNet-derived CNNs

[1] [4]

Evaluation

Upstream reports PULP-Dronet v3 at 19 fps (320 kB) and Tiny-PULP-Dronet v3 at 139 fps (2.9 kB) on the GAP8; testing script reports collision accuracy and yaw-rate RMSE.

Reported figures, not reproduced.

[4]

Project notes

Which version runs on which hardware

Version 1 ran on the group's own PULP-Shield board on a Crazyflie 2.0. Versions 2 and 3 target the Bitcraze AI-deck on a Crazyflie 2.1, which carries a GAP8 processor and a Himax camera. This profile pins the pulp-dronet-v3 release. It ships two pre-trained PyTorch models: PULP-Dronet v3, reported at 19 frames per second and 320 kB, and Tiny-PULP-Dronet v3, reported at 139 frames per second and 2.9 kB on the GAP8. The testing script reports collision accuracy and yaw-rate error. Those figures are upstream reports, not results reproduced here. [1] [4] [6]

What a builder must match

Deployment is tied to specific versions. The v3 instructions require gap-sdk 3.9.1, crazyflie-firmware tag 2022.01, the Crazyflie client at 2022.01 and the Python library at 0.1.17.1. They warn that newer Crazyflie firmware is not compatible because it changed substantially. A newly bought AI-deck may also need specific NINA firmware, and flashing the GAP8 uses a JTAG cable. The STM32 flight application then reads the network's outputs over UART and starts flying when the fly parameter is set in the client. Anyone using the catalog's Crazyflie profile should compare its firmware revision against these pins before assuming the two work together. [4]

The archive does not contain the submodules for the NEMO and DORY tools, the viewer or the modified firmware. Those folders are empty, so a recursive clone is required, and several submodule URLs use SSH addresses. Training uses a conda environment and GPU-flagged scripts. [4] [6]

What we checked

Source research only; no training, installation, flashing or flight test performed. No affiliate links added.

Not yet established

  • Requirements: operating system, host computer

Updates

Upstream last updated
Jul 3, 2025 [2]
Latest release
Mar 10, 2025 [2]
Profile updated
Oct 5, 2026
Sources last checked
Oct 5, 2026
Next review
Nov 5, 2026
Reviewed revision
b68ad38670ba150f7b502e5a848c30d0991beb42 (tag pulp-dronet-v3, release published 2025-03-10) · combined hardware and software repository

Source files and licenses

Code and weights are Apache-2.0; the v3 dataset is noncommercial and external, and deployment depends on pinned external firmware and SDK versions.

LayerAvailability / license
softwareavailable · Apache-2.0
Repository files Apache-2.0 per LICENSE and LICENSE_README.md. External crazyflie-firmware and viewer submodules are GPL-3.0 and not included in the archive. [3] [7]
model weightsavailable · Apache-2.0
Pre-trained .pth files are in the repository and fall under its Apache-2.0 statement; training-data terms for the v3 dataset (CC BY-NC-SA 4.0) are separate and not resolved here. [3] [4] [6]
datasetsavailable · CC BY-NC-SA 4.0
v3 dataset hosted on Zenodo under CC BY-NC-SA 4.0 per upstream README; the Zenodo record was not preserved (HTTP 403). v1 Udacity/Zurich subsets are MIT and live in submodules. [1] [3] [4]
documentationavailable · Apache-2.0
READMEs in the repository under its Apache-2.0 statement. [1] [3]

Dataset and licenses

The v3 dataset was collected with a human pilot flying a Crazyflie 2.1. Images were labelled for collision whenever an obstacle in view was closer than 2 m, and also carry the pilot's yaw-rate commands. The collection framework uses the AI-deck and a Flow deck v2, with an optional Multi-ranger deck. The image counts differ between documents: the main README says 66k labelled images, while the v3 README describes 77k images and tabulates 77,418. [1] [4] [5]

Repository code, documentation and the bundled weights are released under Apache-2.0. The v3 dataset is hosted on Zenodo under CC BY-NC-SA 4.0, which does not grant commercial-use rights. External Crazyflie firmware and viewer modules are GPL-3.0, and the v1 Udacity and Zurich datasets are MIT. Whether noncommercial dataset terms affect use of weights trained on that data was not resolved here. The Zenodo page itself was not preserved. No training, flashing or flight was performed for this profile. [1] [3] [7]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history

2026-10-05 — Initial intake from tag pulp-dronet-v3 (Growbotics comparison lead verified against upstream).