Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs via a Streaming Causal Transformer
Yang Liu · Yifan He · Wenhao Zhao · Xiangyu Mo · Yang Xu · Hao Wei
Mingze Ma · Huan Li · Yifan Wu · Fei Gao · Zipeng Dai · Xin Zhou
arXiv 2026
TIO-Former is a camera-free, optical-flow-free, and mapless range-inertial odometry framework for resource-constrained nano-UAVs. Given six orthogonal 8 × 8 time-of-flight (ToF) arrays and synchronized IMU measurements, it estimates continuous 6-DoF motion with bounded inference cost and memory.
| 384 ToF ranges per frame |
15 g ToF sensor payload |
15 Hz online inference |
10.466 ms P95 edge latency |
| Reliability-aware ToF encoding Bilateral gated differences preserve valid metric changes while suppressing artifacts from invalid returns. |
IMU-guided directional fusion Inertial queries dynamically route the six directional ToF tokens according to platform motion. |
Bounded streaming memory Local KV and compressed Chunk-FIFO memory keep the per-step cost fixed over long flights. |
Architecture of TIO-Former.
- [Sep 15, 2026] The paper is available on arXiv.
- [Sep 14, 2026] The repository was created. Code, pretrained models, and the dataset will be released here.
| Resource | Status |
|---|---|
| Paper | Available on arXiv |
| Training and evaluation code | Coming soon |
| Pretrained models and configurations | Coming soon |
| Dataset and preparation tools | Coming soon |
This work was conducted during the author's internship at Differential Robotics, with Prof. Fei Gao as academic advisor and direct mentorship from Zipeng Dai.
If you find this work useful, please cite:
@misc{liu2026tioformer,
title = {TIO-Former: Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs via a Streaming Causal Transformer},
author = {Liu, Yang and He, Yifan and Zhao, Wenhao and Mo, Xiangyu and Xu, Yang and Wei, Hao and Ma, Mingze and Li, Huan and Wu, Yifan and Gao, Fei and Dai, Zipeng and Zhou, Xin},
year = {2026},
eprint = {2609.17198},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.17198}
}For questions about the paper or future releases, please open an issue in this repository.