Tiny-DroNeRF: Tiny Neural Radiance Fields aboard Federated Learning-enabled Nano-drones

Ilenia Carboni; Elia Cereda; Lorenzo Lamberti; Daniele Malpetti; Francesco Conti; Daniele Palossi

cs.ROMar 2, 2026

Tiny-DroNeRF: Tiny Neural Radiance Fields aboard Federated Learning-enabled Nano-drones

Authors: Ilenia Carboni, Elia Cereda, Lorenzo Lamberti, Daniele Malpetti, Francesco Conti, Daniele Palossi

Organizations: Dalle Molle Institute for Artificial Intelligence, USI-SUPSI, Switzerland. · Department of Electrical, Electronic and Information Engineering, University of Bologna, Italy. · Integrated Systems Laboratory, ETH Zürich, Switzerland.

Abstract

Sub-30g nano-sized aerial robots can leverage their agility and form factor to autonomously explore cluttered and narrow environments, like in industrial inspection and search and rescue missions. However, the price for their tiny size is a strong limit in their resources, i.e., sub-100 mW microcontroller units (MCUs) delivering $\sim$ 100 GOps/s at best, and memory budgets well below 100 MB. Despite these strict constraints, we aim to enable complex vision-based tasks aboard nano-drones, such as dense 3D scene reconstruction: a key robotic task underlying fundamental capabilities like spatial awareness and motion planning. Top-performing 3D reconstruction methods leverage neural radiance fields (NeRF) models, which require GBs of memory and massive computation, usually delivered by high-end GPUs consuming 100s of Watts. Our work introduces Tiny-DroNeRF, a lightweight NeRF model, based on Instant-NGP, and optimized for running on a GAP9 ultra-low-power (ULP) MCU aboard our nano-drones. Then, we further empower our Tiny-DroNeRF by leveraging a collaborative federated learning scheme, which distributes the model training among multiple nano-drones. Our experimental results show a 96% reduction in Tiny-DroNeRF's memory footprint compared to Instant-NGP, with only a 5.7 dB drop in reconstruction accuracy. Finally, our federated learning scheme allows Tiny-DroNeRF to train with an amount of data otherwise impossible to keep in a single drone's memory, increasing the overall reconstruction accuracy. Ultimately, our work combines, for the first time, NeRF training on an ULP MCU with federated learning on nano-drones.

Figures & tables

Method	Year	Neural	Key contributions	PSNR [dB]	SSIM [-]	Params [ $\text{\,}\mathrm{MB}$ ]	Training time	Power class [ $\text{\,}\mathrm{W}$ ]	Device
NeRF [ 1 ]	2021	✓	Coordinate-based MLP	31.01	0.947	5	24 h	$10^{2}$	GPU
SNeRG [ 2 ]	2021	✓	Neural grid	30.38	0.950	1772	15 h	$10^{2}$	GPU
PlenOctrees [ 3 ]	2021	✓	Octree + Spherical harmonics	31.71	0.958	1930	15 h	$10^{2}$	GPU
Plenoxels [ 4 ]	2022	$\times$	Voxel grid + neural network-free	31.71	0.958	778	11 min	$10^{2}$	GPU
Instant-NGP [ 5 ]	2022	✓	Parametric hash-grid	33.18	0.959	46	5 min	$10^{2}$	GPU
TensoRF [ 6 ]	2022	$\times$	Voxel grid w/ tensor factorization	33.14	0.963	72	17 min	$10^{2}$	GPU

TABLE I : Comparison of novel view synthesis methods on the Synthetic NeRF 360° dataset [ 1 ] .

Fig. 1 : A) Novel view synthesis with neural radiance fields and volumetric rendering [ 1 ] . B) The InstantNGP algorithm.

Fig. 2 : Target robot platform. A) Bitcraze Crazyflie 2.1 Brushless with GAP9Shield expansion deck. B) GWT GAP9 System-on-Chip architecture.

Fig. 3 : Memory vs. reconstruction quality Pareto curve. Each point corresponds to a configuration of batch size $B$ (marker) and encoding size $\log_{2}(T)$ (number). Circled in red, the configuration selected for deployment.

Img / step	Res $[px]$	$B$	$\log_{2}(T)$	Mem $[$ \mathrm{MB} $]$	Compute $[$ \mathrm{GOp} $]$	PSNR $[$ \mathrm{dB} $]$	SSIM $[-]$
100	800	264k	19	527	17.52	32.59	0.96
1	160	32k	14	43	2.64	27.19	0.89
			13	39	2.62	27.18	0.90
			12	37	2.61	27.03	0.89
			11	36	2.61	26.61	0.89
1	160	16k	14	31	1.65	27.17	0.89

TABLE II : Results of the hyperparameter optimization starting from the Instant-NGP model [ 5 ] . In bold, the configuration selected for deployment.

Fig. 4 : Kernel latencies and peak on-chip memory usage per training step over a 1024-sample batch tile.

Fig. 5 : A) In-field data collection setup. B) Samples of ground-truth images and the respective reconstructions.

Fig. 6 : A-C) 2D top-view of the IID vs. non-IID dataset distributions, B-D) Federated learning PSNR testing curves for the IID vs. non-IID dataset trainings.

Neural Radiance Fields

Neural radiance fields represent 3D scenes as continuous functions of position and viewing direction to synthesize images from new camera views.

See all 82 papers

Event-Aided Sharp Radiance Field Reconstruction for Fast-Flying Drones
Feb 24, 2026Rong Zou, Marco Cannici, Davide ScaramuzzaNeural Radiance Fields Aerial Robotics
Robotics and Perception Group, University of Zurich, Switzerland
Improving Autonomous Nano-drones Performance via Automated End-to-End Optimization and Deployment of DNNs
Jul 14, 2026Vlad Niculescu, Lorenzo Lamberti, Francesco Conti +2Efficient Neural Network Inference Aerial Robotics
Integrated Systems Laboratory - ETH Zürich, Switzerland · Department of Electrical, Electronic and Information Engineering - University of Bologna, Italy · Dalle Molle Institute for Artificial Intelligence - USI-SUPSI, Switzerland
Neural 3D Object Reconstruction with Small-Scale Unmanned Aerial Vehicles
Sep 15, 2025Àlmos Veres-Vitàlyos, Filip Lemic, Daniel Johannes Bugelnig +5Aerial Robotics 3D Reconstruction
i2CAT Foundation, Spain · Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia · University of Klagenfurt, Austria +3

Feb 24, 2026J/K move · Enter open · S save

Event-Aided Sharp Radiance Field Reconstruction for Fast-Flying Drones

Rong Zou, Marco Cannici, Davide Scaramuzza

Robotics and Perception Group, University of Zurich, Switzerland

Neural Radiance Fields Aerial Robotics Visual-Inertial Odometry

Open paper AskPDF arXiv

Fast-flying aerial robots promise rapid inspection under limited battery constraints, with direct applications in infrastructure inspection, terrain exploration, and search and rescue. However, high speeds lead to severe motion blur in images and induce significant drift and noise in pose estimates, making dense 3D reconstruction with Neural Radiance Fields (NeRFs) particularly challenging due to their high sensitivity to such degradations. In this work, we present a unified framework that leverages asynchronous event streams alongside motion-blurred frames to reconstruct high-fidelity radiance fields from agile drone flights. By embedding event-image fusion into NeRF optimization and jointly refining event-based visual-inertial odometry priors using both event and frame modalities, our method recovers sharp radiance fields and accurate camera trajectories without ground-truth supervision. We validate our approach on both synthetic data and real-world sequences captured by a fast-flying drone. Despite highly dynamic drone flights, where RGB frames are severely degraded by motion blur and pose priors become unreliable, our method reconstructs high-fidelity radiance fields and preserves fine scene details, delivering a performance gain of over 50% on real-world data compared to state-of-the-art methods.

Abstract

Neural Radiance Fields

Explore similar work

Event-Aided Sharp Radiance Field Reconstruction for Fast-Flying Drones

Improving Autonomous Nano-drones Performance via Automated End-to-End Optimization and Deployment of DNNs

Neural 3D Object Reconstruction with Small-Scale Unmanned Aerial Vehicles