Open Source RoboticsRobots, software & learning

PULP-Dronet

Robotics model

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]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Embedded-AI and nano-drone researchers able to train PyTorch models and flash Crazyflie and GAP8 firmware [4]
Parameters
320 kB quantized (~320k params [4] [8]
Embodiments
Bitcraze Crazyflie 2.1 with AI-deck (GAP8 + Himax camera) for v2/v3, Bitcraze Crazyflie 2.0 with PULP-Shield for v1 [1] [4]
Source rights
Named licenses are documented by part. Includes noncommercial terms. Review licenses

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 family of DroNet-derived 2D convolutional neural networks with training, NEMO INT8 quantization, and DORY deployment code for autonomous visual navigation aboard a pocket-size quadrotor. At pinned release pulp-dronet-v3 (b68ad38670ba150f7b502e5a848c30d0991beb42), the repository targets the Bitcraze Crazyflie 2.1 equipped with an AI-deck (GreenWaves GAP8 RISC-V SoC and Himax grayscale camera); v1 targeted the PULP-Shield on Crazyflie 2.0. [1] [2] [4]

Model card summary

Architecture

DroNet-derived 2D CNN family for ultra-low-power microcontrollers: PULP-Dronet v3 (ResBlock with bypass, depth multiplier 1.0) and Tiny-PULP-Dronet v3 (depthwise/pointwise separable convolutions without bypass, depth multiplier 0.125), quantized via NEMO and compiled via DORY for the GAP8 SoC.

[1] [4]

Base model

DroNet (University of Zurich)

[1]

Parameter count

320 kB quantized (~320k params; 1.30 MB FP32 .pth) for PULP-Dronet v3 ResBlock 1.0, and 2.9 kB quantized (~2.9k params; 24 KB FP32 .pth) for Tiny-PULP-Dronet v3 dw-pw 0.125

[4] [8]

Input modalities
rgb

Monocular grayscale Himax camera images on the Bitcraze AI-deck (mapped to closest schema enum rgb).

[4] · Editorial assessment

Output modalities
waypoints

Outputs scalar collision probability (obstacle within 2 m) and continuous steering yaw-rate command over UART to the Crazyflie flight controller (mapped to closest schema enum waypoints).

[4] · Editorial assessment

Action space

Two scalar visual-navigation outputs per grayscale camera frame: collision probability (obstacle-within-2m avoidance trigger) and steering yaw-rate command, passed over UART from GAP8 to the Crazyflie STM32 flight application.

[4]

Control frequency

19 Hz (PULP-Dronet v3) and 139 Hz (Tiny-PULP-Dronet v3) upstream-reported onboard inference rates on GAP8

[4]

Trained / supported embodiments
  • Bitcraze Crazyflie 2.1 with AI-deck (GAP8 + Himax camera) for v2/v3
  • Bitcraze Crazyflie 2.0 with PULP-Shield for v1

[1] [4]

Training datasets
  • PULP-Dronet v3 Crazyflie dataset on Zenodo (66k–77,418 labelled grayscale flight images, CC BY-NC-SA 4.0); v1 Udacity and Zurich datasets (MIT) in submodules

[1] [3] [4]

Interfaces
  • UART (GAP8 AI-deck to Crazyflie STM32)
  • PyTorch .pth / NEMO ONNX / DORY C codegen
  • cfclient fly parameter

[4]

Checkpoints

CheckpointFormatSizeAccessPinned revision
pulp-dronet-v3-resblock-1.0.pthpytorch0.0013 GBopenb68ad38670ba
tiny-pulp-dronet-v3-dw-pw-0.125.pthpytorch0.000024 GBopenb68ad38670ba

Pretrained FP32 .pth weights committed directly in the repository at tag pulp-dronet-v3 (1,298,645 bytes and 24,085 bytes).

[4] [6] [8]

Is this for you?

Use cases
  • Onboard collision avoidance and steering on nano-drones
  • Dataset collection on a nano-drone
  • Study of CNN quantization and deployment on microcontrollers

[1] [4]

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

Safety and limitations
  • Onboard deployment is tightly pinned to crazyflie-firmware tag 2022.01, gap-sdk 3.9.1, crazyflie-clients-python 2022.01, and crazyflie-lib-python 0.1.17.1; upstream warns newer firmware versions are incompatible.
  • v3 training dataset on Zenodo is licensed CC BY-NC-SA 4.0 (noncommercial); whether noncommercial dataset terms affect weights trained on it is unresolved here.
  • Submodules (NEMO, DORY, viewer, modified firmware) require recursive git clone and include GPL-3.0 external components.
  • No training, quantization, firmware flashing, or flight testing was performed by Open Source Robotics.

[3] [4] [6]

Compute & fine-tuning

Inference compute
  • Minimum VRAM: 0 GB
  • Documented hardware: GreenWaves GAP8 RISC-V SoC on Bitcraze AI-deck (quantized INT8 onboard; no GPU required at inference)
  • Precision: int8
  • Edge / consumer capable: Yes

Quantized via NEMO and deployed via DORY to run fully aboard the GAP8 microcontroller.

[4]

Fine-tuning

[4]

Host computer

GPU host for PyTorch training (--gpu flag); onboard inference runs quantized INT8 directly on the GreenWaves GAP8 microcontroller aboard the Bitcraze AI-deck (JTAG cable required to flash GAP8).

[4]

Operating system
  • Linux

Inferred from gap-sdk 3.9.1, conda_deps.yml, and Crazyflie build toolchain.

[4] · 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]

Setup & usage

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

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
PULP-Dronet v3 GAP8 onboard throughput and test-set evaluation (collision accuracy & yaw-rate RMSE)PULP-Dronet v3: 19 fps (320 kB); Tiny-PULP-Dronet v3: 139 fps (2.9 kB) on GAP8Upstream-reported throughput figures in tiny-pulp-dronet-v3/README.md; OSR has not trained, quantized, flashed, or flown the models.

[4]

Project notes

Models, checkpoints, and onboard action interface

The v3 release ships two pretrained PyTorch checkpoints directly in tiny-pulp-dronet-v3/model/: pulp-dronet-v3-resblock-1.0.pth (1,298,645 bytes / 1.30 MB FP32; ResBlock with bypass, depth multiplier 1.0, 320 kB quantized) and tiny-pulp-dronet-v3-dw-pw-0.125.pth (24,085 bytes / 24 KB FP32; depthwise-pointwise separable convolutions without bypass, depth multiplier 0.125, 2.9 kB quantized). Both take monocular grayscale Himax camera frames and output two scalar values—collision probability (obstacle within 2 m) and steering yaw-rate—sent over UART from the GAP8 to the Crazyflie STM32 flight application. [4] [6] [8]

Deployment is strictly pinned to 2022 toolchains: gap-sdk 3.9.1, crazyflie-firmware tag 2022.01, crazyflie-clients-python 2022.01, and crazyflie-lib-python 0.1.17.1. Upstream warns that newer Crazyflie firmware is incompatible, newly bought AI-decks may need specific NINA firmware, flashing GAP8 requires a JTAG cable, and recursive cloning is needed to populate empty submodule paths (NEMO, DORY, viewer, and modified firmware). [4] [6]

Upstream-reported throughput and unverified execution

The v3 README reports onboard GAP8 inference at 19 fps (320 kB) for PULP-Dronet v3 and 139 fps (2.9 kB) for Tiny-PULP-Dronet v3, with testing.py evaluating collision accuracy and yaw-rate RMSE. All performance figures are upstream-reported; Open Source Robotics has not run training, quantization, inference, flashing, or flight tests. [4]

What we checked

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

Updates

Upstream last updated
Jul 3, 2025 [2]
Latest release
Mar 10, 2025 [2]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
b68ad38670ba150f7b502e5a848c30d0991beb42 (tag pulp-dronet-v3, release published 2025-03-10) · combined hardware and software repository

Separate license matrix

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
Codeavailable · 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]
Training 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]

Split code, weight, and dataset licenses

Repository code, documentation, and bundled .pth weights carry an Apache-2.0 license. However, the v3 flight dataset (reported as 66k images in the main README and 77,418 tabulated in the v3 README) is hosted on Zenodo under CC BY-NC-SA 4.0 (noncommercial), external Crazyflie firmware and viewer submodules are GPL-3.0, and v1 Udacity/Zurich subsets are MIT. Whether noncommercial dataset terms affect downstream reuse of weights trained on that data is not resolved here. [1] [3] [4] [5] [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 0.8.0-draft.1

Change history

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

2026-11-21 — Migrated from 0.7.0-draft.5 project to 0.8.0-draft.1 model; added verified .pth checkpoint entries (1.30 MB and 24 KB), GAP8 INT8 edge compute profile, and upstream-reported throughput benchmarks.