AirIMU : Learning Uncertainty Propagation for Inertial Odometry

📢 Latest News

• [2025-02-01] Introducing Our New Project!<br> 🚀 AirIO : Learning Inertial Odometry with Enhanced IMU Feature Observability<br>

AirIO achieves up to 86.6% performance boost over SOTA methods:

- ✅ Tailored specifically for drones
- ✅ No external sensors or control inputs required
- ✅ Generalizes to unseen trajectories
- ✅ Explicitly encodes UAV attitude and predicts velocity in body-frame representation

Installation

This work is based on pypose. Follow the instruction and install the newest realase of pypose: https://github.com/pypose/pypose

Dataset

Note: Remember to reset the data_root in configs/datasets/${DATASET}/${DATASET}.conf.

Download the Euroc dataset from: https://projects.asl.ethz.ch/datasets/doku.php?id=kmavvisualinertialdatasets

Download the TUM VI dataset from: https://cvg.cit.tum.de/data/datasets/visual-inertial-dataset

Download the KITTI dataset from: https://www.cvlibs.net/datasets/kitti/

Download the SubT-MRS dataset: Download. 🎈You can find more information about SubT-MRS datasets from the links here.

Pretrained Model

Note: You can download our trained ckpt here.

KITTI

EuRoC

Train

Easy way to start the training using the exisiting configuration.

Note:You can also create your own configuration file for different datasets and set the parameters accordingly.

python train.py --config configs/exp/EuRoC/codenet.conf

More specific option:

usage: train.py [-h] [--config CONFIG] [--device DEVICE] [--load_ckpt] [--log]

optional arguments:
  -h, --help       show this help message and exit
  --config CONFIG  config file path
  --device DEVICE  cuda or cpu, Default is cuda:0
  --load_ckpt      If True, try to load the newest.ckpt in the exp_dir specificed in our config file.
  --log            if True, save the meta data with wandb, Default is True

Evaluation

To evaluate the model and generate network inference file net_output.pickle, run the following command:

python inference.py --config configs/exp/EuRoC/codenet.conf

Inference on a standalone IMU CSV

If only the raw IMU measurements are available (timestamps, gyroscope and accelerometer) you can run the network directly on the CSV files listed in the dataset configuration without any ground‑truth orientation:

python inference_offline.py --config configs/exp/EuRoC/codenet.conf \
                        --ckpt experiments/EuRoC/codenet/ckpt/best_model.ckpt

The script reads each sequence's mav0/imu0/data.csv based on data_root and data_drive entries in the config and writes net_output.pickle keyed by sequence name containing corrected IMU data (corrected_acc, corrected_gyro and dt).

Exporting to ONNX

Convert a training checkpoint to both ONNX and a plain PyTorch state dictionary:

python export_onnx.py --config configs/exp/EuRoC/codenet.conf \
                      --ckpt experiments/EuRoC/codenet/ckpt/best_model.ckpt \
                      --onnx airimu_euroc.onnx --torch airimu_euroc.pt

By default the ONNX model is exported in float32 precision to ensure compatibility with the CPU execution provider of ONNX Runtime. If you need a float64 graph, pass --fp64 (note that some runtimes lack double-precision Conv support).

Offline inference with ONNX

Run an exported ONNX model directly on the IMU CSV files listed in the config:

python inference_onnx.py --config configs/exp/EuRoC/codenet.conf \
                         --onnx airimu_euroc.onnx --seqlen 1000

Pass --whole to process an entire sequence in a single pass; otherwise the IMU stream is chunked into windows of length --seqlen, mirroring the behaviour of inference.py.

The resulting net_output.pickle stores all arrays as PyTorch tensors, matching the format produced by inference_offline.py so that downstream evaluation scripts that call .cpu() work without modification.

You can use the evaluation tool to assess your model performance with net_output.pickle， run the following command.

Note: Make sure to replace path/to/net_output_directory with the directory path where your network output pickle file is stored.

python evaluation/evaluate_state.py --dataconf configs/datasets/${DATASET}/${DATASET}.conf --exp path/to/net_output_directory

More specific option for the evaluation tool:

usage: evaluation/evaluate_state.py [-h] [--dataconf] [--device] [--exp] [--seqlen] [--savedir] [--usegtrot] [--mask]

optional arguments:
  -h, --help       show this help message and exit
  --config         config file path
  --device         cuda or cpu, Default is cuda:0
  --exp            the directory path where your network output pickle file is stored
  --seqlen         the length of the integration sequence
  --savedir        the save diretory for the evaluation results, default path is  "./result/loss_result"
  --usegtrot       use ground truth rotation for gravity compensation, default is true
  --mask           mask the segments if needed. 

Cite Our Work

Thanks for using our work. You can cite it as:

@article{qiu2023airimu,
            title={AirIMU: Learning Uncertainty Propagation for Inertial Odometry}, 
            author={Yuheng Qiu and Chen Wang and Can Xu and Yutian Chen and Xunfei Zhou and Youjie Xia and Sebastian Scherer},
            year={2023},
            eprint={2310.04874},
            archivePrefix={arXiv},
            primaryClass={cs.RO}
      }