TinyCVIO: A Constellation-Aided Visual-Inertial Odometry System for Nanodrones

Derin Ozturk; Kaan Akan; Irwin Wang; Christopher Batten

cs.ROSep 24, 2026

TinyCVIO: A Constellation-Aided Visual-Inertial Odometry System for Nanodrones

Authors: Derin Ozturk, Kaan Akan, Irwin Wang, Christopher Batten

Organizations: Department of ECE, Cornell University, Ithaca, NY, USA

Abstract

Nanodrones require accurate, real-time state estimation under severe sensing and computational constraints. We present TinyCVIO, a visual-inertial odometry system that co-designs miniature sensing, visual processing, and estimation for a commodity dual-core microcontroller with 520 kB SRAM. Lightweight LED constellations provide known geometry without surveyed positions or yaw angles, assuming placement on a common level plane. A streaming visual frontend tracks LED observations from a millimeter-scale camera at 29.2 FPS, while a rigid-board measurement model retains inter-LED constraints and streaming QR bounds estimation workspace for a fixed filter-state size. Across 19 hand-held hardware-in-the-loop datasets, the rigid-board model reduces mean absolute trajectory error by 27% relative to planar points. The complete system runs onboard a Crazyflie across nine flights at three speeds, achieving 3.5-3.7 cm mean absolute trajectory error and 0.50-0.60% relative pose error over 10 m segments, with mean estimate latency of 15.7-16.3 ms.

Figures & tables

Fig. 1: TinyCVIO hardware and estimation concept. Left: Crazyflie carrying our custom TinyCVIO deck, with the NanEyeC camera connected through a custom flexible printed circuit board cable and the ICM-45686 IMU indicated. Right, top to bottom: observations of shared LED constellations across successive camera poses; RP2354-based companion compute deck for a Crazyflie; and five-LED constellation with known geometry.

Fig. 2: RP2354 architecture and reserved memory allocation. The lower map omits Core 0 stack allocation in SRAM8–9.

Fig. 3: Visual frontend pipeline and example observations. (a) Spatial grouping and connected components, with an enlarged constellation. (b) Geometry-based LED identification, including candidate separation of a merged component. (c) Tracking preserves LED identities between frames using a fitted similarity transform; open circles denote previous positions and filled circles current tracked positions. Outputs contain image coordinates and board and LED identifiers.

Model	Tracks	DOF/track	Total fold rows
Rigid board	$1$	$3$	$qF+3(F-1)$
Planar points	$N_{\ell}$	$2$	$N_{\ell}(2F-2)$
Free points	$N_{\ell}$	$3$	$N_{\ell}(2F-3)$

TABLE I: Processing one board’s

N_{\ell}

LEDs over

F

frames.

Fig. 4: Frontend performance over six hand-held captures, binned by angular velocity, board distance, and image radius. Rows show localization error (mean, RMS, 5–95% band), mislabeled detections by type (dashed: aggregate), in-view boards excluding the tracker’s border exclusion zone that are detected (solid) and accepted by the backend (dashed), and board counts.

Fig. 5: Median ATE over

350

runs per point, all axes log-scaled: (a) detection noise, with interquartile bands; (b) per-update LED budget at

0.5\text{\,}\mathrm{p}\mathrm{x}

(solid) and

2\text{\,}\mathrm{p}\mathrm{x}

(dashed) noise. Board budgets increase in steps of five LEDs. Free points diverge in 3–15% of runs; neither constrained model diverges.

	Rigid board			Planar point
Capture	B1	B2	B3	P1	P2	P3
Path (m)	39.8	45.0	41.8	52.3	53.1	30.5
Median speed (m/s)	0.77	0.87	0.81	0.98	1.04	0.47
ATE (cm)	1.92	3.43	1.75	2.72	3.61	2.81
Orientation ( ∘ )	1.13	1.35	1.63	1.84	2.39	1.17
RPE ${}_{10\,\mathrm{m}}$ (%)	0.29	0.38	0.26	0.37	0.43	0.39

TABLE II: Live HITL estimation accuracy and timing.

Model	ATE (cm)	Ori. ( ∘ )	RPE ${}_{10\,\mathrm{m}}$ (%)
Rigid board	2.44	1.91	0.34
Planar point	3.32	2.38	0.46

TABLE III: Mean errors over 19 paired HITL capture replays.

Speed (m/s)	0.34	0.41	0.50
ATE (cm)	3.65 $\pm$ 0.43	3.48 $\pm$ 0.34	3.71 $\pm$ 0.79
Ori. ( ∘ )	1.44 $\pm$ 0.31	1.31 $\pm$ 0.68	1.20 $\pm$ 0.26
RPE ${}_{10\,\mathrm{m}}$ (%)	0.57 $\pm$ 0.06	0.50 $\pm$ 0.10	0.60 $\pm$ 0.15
Mean est. latency (ms)	15.7 $\pm$ 0.1	15.7 $\pm$ 0.1	16.3 $\pm$ 0.1
Peak est. latency (ms)	30.7	31.9	31.2

TABLE IV: Flight results across three trials per speed.

Fig. 6: Estimated and Vicon trajectories for one flight in the fastest speed group, after SE(3) alignment.

Visual-Inertial Odometry

Visual-inertial odometry estimates camera motion by combining image geometry with inertial measurements for metric navigation.

See all 54 papers

TIO-Former: Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs via a Streaming Causal Transformer
Sep 15, 2026Yang Liu, Yifan He, Wenhao Zhao +9Inertial Odometry Transformer
Zhejiang University, Hangzhou, China · Differential Robotics, Hangzhou, China
Minimalist Visual Inertial Odometry
May 19, 2026Francesco Pasti, Jeremy Klotz, Nicola Bellotto +1Inertial Odometry Mobile Robotics
Department of Information Engineering, University of Padua, Padua, Italy · Computer Science Department, Columbia University, New York, NY, USA
Borrowed Eyes: Markerless Nano-UAV Flight with an Active Quadruped Observer
Oct 7, 2026Alejandro Lorite Mora, Dimitrios Arapis, Andrés FaíñaUAV Navigation GNSS-Denied Localization
IT University of Copenhagen, Denmark · Novo Nordisk A/S, Hillerød, Denmark

Sep 15, 2026J/K move · Enter open · S save

TIO-Former: Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs via a Streaming Causal Transformer

Yang Liu, Yifan He, Wenhao Zhao, Xiangyu Mo+8

Zhejiang University, Hangzhou, China · Differential Robotics, Hangzhou, China

Inertial Odometry Transformer UAV Navigation Robot State Estimation Multimodal Sensor Fusion

Open paper AskPDF arXiv

Autonomous nano-UAV navigation requires accurate ego-motion estimation under stringent size, weight, power, and computing (SWaP-C) constraints, where visual sensors and LiDARs exceed payload limits, optical flow degrades in low-texture scenes, and inertial-only state estimation is susceptible to accumulated drift. While multi-zone time-of-flight (ToF) arrays provide a lightweight metric complement, 6-DoF estimation from merely 384 ranges per frame is challenged by invalid returns, anisotropic observability, and temporal computational scaling. We propose TIO-FORMER, a camera-free, optical-flow-free, and mapless range-inertial odometry framework driven by an IMU and an ultra-lightweight (15 g) payload of six orthogonal 8 x 8 ToF arrays. Our frontend pairs consecutive range grids with a bilateral gated difference, while IMU-guided cross-attention dynamically routes directional features conditioned on platform kinematics. A Streaming Causal Transformer couples an uncompressed Local KV cache with compressed Chunk-FIFO memory, maintaining bounded inference cost and memory footprint independent of flight duration. In real-flight evaluations, TIO-FORMER reduces open-loop position error by 54.4% compared to nano-UAV optical flow and by 66.4%-89.1% over learned inertial baselines. We also evaluate performance across multiple environments and robustness under severe sensing degradation. Deployed on an edge RISC-V companion computer, TIO-FORMER achieves a P95 latency of 10.466 ms and peak resident memory of 6.324 MiB (less than 5 percent system RAM), demonstrating that sparse range sensing provides practical geometric anchoring for resource-constrained micro-aerial robots. Code is available at https://github.com/Ly041021/TIO-Former.

Abstract

Visual-Inertial Odometry

Explore similar work

TIO-Former: Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs via a Streaming Causal Transformer

Minimalist Visual Inertial Odometry

Borrowed Eyes: Markerless Nano-UAV Flight with an Active Quadruped Observer