Research map / Label-efficient and robust learning / Test-time adaptation
Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Tartan IMU, a multi-platform pretrained IMU odometry foundation model with LoRA adaptation and online test-time adaptation via GMM-based buffer selection and SLAM supervision.
Problem
Learning-based inertial odometry generalizes poorly across robotic platforms and motion patterns, overfits to specific datasets, and struggles to adapt to new platforms.
Approach
A ResNet+LSTM shared IMU backbone with robot-specific regression heads, pretrained on over 100 hours of multi-platform IMU data in the body frame, then adapted with LoRA and updated online using SLAM-supervised relative poses and a GMM-based balanced training buffer.
Key results
- Tartan IMU achieves on average 35.5% ATE and 41.0% T-RTE improvement over the second-best specialized model across wheeled, handheld, legged, and aerial platforms.
- LoRA fine-tuning with 1.1M trainable parameters adapts to unseen TartanDrive data with 33% fewer iterations than the SOTA RNIN-VIO baseline.
- Online adaptation completes adaptation from SubT UGV model to TartanDrive car model within 105 seconds, running at 200 FPS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the shared model on data from all robot platforms instead of a single platform (ablation) | Tartan IMU · Wheeled, handled, legged, and aerial platforms | ATE: Wheeled: 7.51 -> 6.17; Handled: 6.64 -> 4.32; Legged: 3.64 -> 1.46; Aerial: 3.84 -> 3.32 | Helped |
| Train the shared model on data from all robot platforms instead of a single platform (ablation) | Tartan IMU · Wheeled, handled, legged, and aerial platforms | T-RTE: Wheeled: 3.17 -> 2.52; Handled: 4.24 -> 1.95; Legged: 1.94 -> 0.79; Aerial: 1.36 -> 1.04 | Helped |
| Use LoRA fine-tuning instead of full fine-tuning (ablation) | Tartan IMU · SubT source domain after fine-tuning on TartanDrive | pose error | Helped |
| Use adaptive training-buffer selection to balance motion-pattern samples (ablation) | Tartan IMU online adaptation · TartanDrive online adaptation | adaptation time | Helped |
| Pretrain on diverse multi-platform data rather than single-platform data (ablation) | Tartan IMU · Wheeled, handled, legged, and aerial platforms | ATE: up to 42% average improvement on ATE | Helped |
| Fine-tune the pretrained model on unseen data | Tartan IMU · TartanDrive | fine-tuning iterations: 33% fewer iterations | Helped |
Surprising findings
- Full fine-tuning causes catastrophic forgetting on the source domain, whereas LoRA preserves source performance.
Limitations
- Cannot support arbitrary robotic platforms without new motion heads.
- GMM-based buffer selection may struggle with complex motion patterns.
- Relies on SLAM as teacher for online adaptation.
Experimental setup
Datasets: SubT-MRS IDOL Blackbird UZH dataset TartanDrive
Benchmarks: SubT-MRS IDOL Blackbird TartanDrive
Metrics: ATE T-RTE
Baselines: AI-IMU RNIN-VIO TLIO IMO IDOL EqNIO
Builds on: TLIO LoRA AI-IMU RIO
Closest papers by meaning
- EqNIO: Subequivariant Neural Inertial Odometry (ICLR 2025 · similarity 0.87)
- Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · similarity 0.85)
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.85)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.84)
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception (CVPR 2024 · similarity 0.83)
- Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation (NeurIPS 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.