MosaicIMU: Composing Carrier Experts for Generalizable Neural Inertial Odometry
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "MosaicIMU: Composing Carrier Experts for Generalizable Neural Inertial Odometry".
Dev: Here's the gist of the paper. Robust inertial odometry is essential for various carriers when external sensing is unreliable Learningbased methods reduce integration drift by capturing local motion priors, but these methods often remain tied to a particular carrier, limiting generalization across heterogeneous platforms We present MosaicIMU, a carrier-conditioned Mixture-of-Experts (MoE) pretraining-and-adaptation framework for generalizable neural inertial odometry MosaicIMU uses a prototype-based router to compose carrier-specific expert features, decodes local velocity and uncertainty constraints, and integrates them with a history-aware EKF Experiments show that MosaicIMU consistently outperforms learning-based baselines, reducing average ATE and RTE10s by 40% and 34%, respectively These results highlight that MosaicIMU provides a scalable pretraining-to-deployment paradigm for generalizable and adaptive neural inertial odometry.
Rosa: Pruning redundant reasoning tokens MosaicIMU is a carrier-conditioned Mixture-of-Experts (MoE) framework designed to achieve sampleadaptive, feature-level expert composition and integrate local velocity constraints with a history-aware EKF for stable state estimation.
Dev: The framework utilizes a prototype-based router to compose carrier-specific expert features and then decodes local velocity and uncertainty constraints before integrating them into the history-aware EKF.
Rosa: For unseen domain adaptation, the method freezes the pretrained base model and learns a new lightweight expert residual branch for efficient adaptation.
Paper summary: Rosa: So, we're wrapping up our look at MosaicIMU today, titled "Composing Carrier Experts for Generalizable Neural Inertial Odometry." Basically, this paper is about how they built a system that can handle different types of sensors—different carriers—without needing a completely new training setup every time.
Dev: Right. It’s taking the idea of using many specialized experts and making them work together intelligently based on what kind of motion you see right now. They used this mixture-of-experts framework to get better results when generalizing across different platforms compared to just training one model for one specific car or robot.
Taro: What really stands out is how they handled adaptation online. They showed that even when the system encounters a completely new type of motion pattern, it can pick informative data points from the real world to update itself efficiently without having to start all over again from scratch.
Rosa: That’s what I liked about it. It moves us toward something that could be truly general purpose neural inertial odometry, not just something tuned perfectly for one specific vehicle or setting. The authors also found that by using a soft way of routing these experts, the system is much more flexible when dealing with different kinds of motion physics.
Dev: Flexibility is good for real-time control. It means the estimation loop doesn't break as easily when things get weird on the road or in a warehouse. They also found that they could achieve a forty percent drop in error metrics like ATE just by making these smart choices about which expert features to combine.
Taro: And those results on vehicle and pedestrian sequences, where the motion is naturally very different, really show that this method works well for heterogeneous motion patterns. It’s not just fine on one thing; it's good at handling the mix of things you see in the real world.
Rosa: Exactly. So, MosaicIMU provides a scalable way to get neural odometry that can adapt to new situations and different hardware better than what we have now. But we've got some caveats, right?
Dev: Yeah, the authors were upfront about it. They said the current performance is still tied to the amount of data they used and which specific carriers they tested on. Also, for real deployment, you still need reasonably stable sensor readings; if things get too chaotic or your mounting changes drastically, it might start failing.
Taro: So while the potential for generalization is there—being able to handle new motion types—we can't assume it will work flawlessly in every single crazy scenario without more robust data and better uncertainty handling.
Rosa: That’s the balance of it. It’s a solid foundation, but we still have some work to do before you can fully trust this as a general solution for every application. Next up, we're going to look at how they actually trained this whole thing—the math behind those expert losses and prototypes.
Conclusion: Rosa: So, to wrap up this discussion on MosaicIMU, we're talking about how they built this system called "Composing Carrier Experts for Generalizable Neural Inertial Odometry."
Dev: Right. This whole framework is essentially about taking those different carrier-specific motion models and stitching them together intelligently so you get one system that works better everywhere than a single model trained for just one type of vehicle.
Taro: What this implies is that we can build odometry that doesn't break when you switch from driving a car to walking down the street, because it learns how to mix those motion types.
Rosa: It's about making neural inertial odometry more robust and adaptable across different hardware and environments without needing a completely fresh training session for every new setup. The authors show they achieved this by using this mixture-of-experts structure that lets the system decide which expert features are most relevant for the current motion situation.
Dev: And they did get solid numbers showing that this approach reduces errors, like the average absolute trajectory error, by forty percent and relative trajectory error by thirty-four percent across four different carrier categories. That’s a big jump for real navigation systems.
Taro: But it's not magic either. They pointed out that the method still depends on how much data you train it on and what specific sensors you tested it with, so we can't assume it will handle every weird motion pattern imaginable without more training.
Rosa: Exactly. It’s a strong step toward generalizable odometry, but we still need to figure out how to make that router even smarter about uncertainty estimation when things get really unexpected out there.
Dev: And that’s what we'll be diving into next—we need to look at the actual math behind how they built these expert losses and prototypes so we can understand exactly how they achieved these results.
Junye Zou, Huiyi Yan, Xinning Xu, Xiaolei Li, Pengkun Zhou, Jinhui Zhang, Ziyang Meng
Tsinghua University
cs.RO
Submitted: 2026-06-08
Updated: 2026-11-06
Comments: Accepted to the Conference on Robot Learning (CoRL 2026) as a Spotlight paper
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 98/100
The gist: Robust inertial odometry is essential for various carriers when external sensing is unreliable Learningbased methods reduce integration drift by capturing local motion priors…
Key concepts
- Mixture-of-Experts (MoE)
- A framework where different specialized neural networks (experts) work together. MosaicIMU uses this by having a router select and combine features from carrier-specific experts based on the current motion, allowing the system to handle different vehicle types effectively.
- Prototype-based Router
- This is a mechanism that decides which expert features are most relevant for a given situation. It compares incoming inertial data against learned 'prototypes' representing different carrier characteristics to generate routing weights, enabling sample-adaptive feature composition.
- History-aware EKF
- An Extended Kalman Filter (EKF) enhanced to consider past poses and local velocity constraints. This allows the system to maintain a stable state estimate by comparing the current prediction with historical trajectory information, ensuring consistency without needing absolute global yaw observations.
Terminology
Summary
Robust inertial odometry is essential for various carriers when external sensing is unreliable Learningbased methods reduce integration drift by capturing local motion priors, but these methods often remain tied to a particular carrier, limiting generalization across heterogeneous platforms We present MosaicIMU, a carrier-conditioned Mixture-of-Experts (MoE) pretraining-and-adaptation framework for generalizable neural inertial odometry MosaicIMU uses a prototype-based router to compose carrier-specific expert features, decodes local velocity and uncertainty constraints, and integrates them with a history-aware EKF Experiments show that MosaicIMU consistently outperforms learning-based baselines, reducing average ATE and RTE10s by 40% and 34%, respectively These results highlight that MosaicIMU provides a scalable pretraining-to-deployment paradigm for generalizable and adaptive neural inertial odometry. p. 0
Pruning redundant reasoning tokens
MosaicIMU is a carrier-conditioned Mixture-of-Experts (MoE) framework designed to achieve sampleadaptive, feature-level expert composition and integrate local velocity constraints with a history-aware EKF for stable state estimation. The framework utilizes a prototype-based router to compose carrier-specific expert features and then decodes local velocity and uncertainty constraints before integrating them into the history-aware EKF. p. 0 For unseen domain adaptation, the method freezes the pretrained base model and learns a new lightweight expert residual branch for efficient adaptation. p. 0 Furthermore, for edge-deployment, it reuses the router to select informative online samples for efficient incremental updates. p. 0
How the training signal is built
The base model is trained with heteroscedastic velocity regression, an auxiliary velocity loss, prototype decorrelation, and a carrier-prior routing regularization. The training objective includes the negative log-likelihood (LNLL), which measures the error between predicted and actual velocity, the auxiliary velocity loss (Lvel), which minimizes the difference between predicted and ground-truth velocities, prototype decorrelation loss (Lorth) to prevent prototype collapse, and a carrier-prior routing regularization term (Lcar) using KL divergence. p. 3
Where the gains come from
The primary gains stem from the sampleadaptive composition of carrier-specific motion priors achieved through the MoE architecture. The router maps encoded inertial features to similarity logits and compares them with learnable carrier prototypes to generate routing weights. p. 3 This allows for the composition of expert features at the feature level before a shared decoder, enabling sample-adaptive composition of carrier-specific motion priors. p. 3 Additionally, segment-level velocity constraints are introduced by aggregating dense local velocities into a segment-level average-velocity observation for the history-aware EKF. p. 3
The filter maintains cloned historical poses and uses a segment aggregation operator to derive the neural observation, which is then compared against the velocity-consistency measurement model to generate residuals This design ensures that the residual enforces consistency between the learned gravity-aligned local velocity and the historical trajectory without introducing an absolute global-yaw observation. The mechanism for online adaptation involves using a mismatch score to select high-mismatch samples from an online buffer to update only lightweight components while keeping the base MosaicIMU parameters frozen. p. 4
The performance gains are evident across different evaluation aspects, such as reducing average ATE and RTE10s by 40% and 34% compared to learning-based baselines. The use of the local gravity-aligned frame L0 for velocity prediction is shown to achieve the best overall performance across all evaluated carriers and metrics, reducing average ATE from 8. p. 700 m to 4.15 m. The router feature and expert-composition analysis reveals that the learned prototypes preserve meaningful carrier semantics while allowing for complementary information sharing among experts. p. 1
The results confirm that the local gravity-aligned frame L0 achieves the best overall performance across all evaluated carriers and metrics, reducing average ATE from 8.00 m to 4.15 m compared with the Bt and W variants. p. 5 The comparison of expert designs shows that MosaicIMU achieves the lowest ATE and RTE-10s across all carriers. p. 6 Soft routing improves the average ATE from 4. p. 771 m to 4.15 m, indicating that soft feature-level composition provides a more flexible mechanism for heterogeneous inertial dynamics than a rigid one-hot expert assignment. p. 6 The router's behavior shows that non-zero off-diagonal weights indicate the router can borrow complementary information from other carrier experts when forming the fused feature.
The final evaluation of online adaptation on a real-world vehicle demonstrates substantial trajectory accuracy improvement, with ATE decreasing to 10.29 m after 10 epochs and further to 7. p. 660 m after 40 epochs. p. 6 The router serves not only as an expert-assignment module but also as a compact signal for selecting informative online samples under edge-computing constraints. p. 7 This demonstrates that the selected samples capture transferable distribution shifts rather than merely overfitting to one trajectory. p. 7 The overall results suggest that MosaicIMU offers a practical and scalable step toward general-purpose neural inertial odometry across heterogeneous carriers. p. 7
Despite the improved cross-carrier performance, MosaicIMU remains limited by the current data scale and carrier types, and thus is not expected to handle all possible IMU motion patterns. The filter update also assumes reasonably stable IMU propagation and attitude estimation, so severe bias changes, unusual sensor mounting, and long-term deployment may lead to potential failure cases. p. 7 Future work will broaden the pretraining data, strengthen router uncertainty estimation, and explore more robust self-supervised online adaptation. p. 7 The paper concludes that MosaicIMU provides a scalable path toward generalizable and adaptive neural inertial odometry. p. 1
MosaicIMU achieves the best average performance and the lowest ATE across all baselines, suggesting that sample-adaptive feature-level composition provides a more flexible inertial prior than either carrier-specific training or fixed-head multi-carrier training. The results establish that MosaicIMU provides a scalable pretraining-to-deployment paradigm for generalizable and adaptive neural inertial odometry. p. 0 p. 0 p. 0 This framework is particularly effective on vehicle and pedestrian sequences, where heterogeneous motion patterns are difficult to capture with a fixed output partition. p. 5 p. 5 p. 5 The gains are especially clear on vehicle and pedestrian sequences, where heterogeneous motion patterns are difficult to capture with a fixed output partition. p. 5 p. 5 p. 5
The comparison of expert designs shows that MosaicIMU achieves the best ATE and RTE-10s across all carriers compared with the singlecarrier expert baseline. The mixture single expert has a higher average ATE of 5. p. 653 m, showing that a single shared expert is insufficient to capture heterogeneous inertial dynamics. p. 6 MosaicIMU combines shared representation learning with carrier-aware expert specialization, verifying the importance of sample-adaptive feature-level composition for multi-carrier inertial learning. p. 6 The results verify the importance of sample-adaptive feature-level composition for multi-carrier inertial learning. p. 6 The overall results show that soft feature-level composition provides a more flexible mechanism for heterogeneous inertial dynamics than a rigid one-hot expert assignment. p. 6 The router's behavior shows that the learned prototypes preserve meaningful carrier semantics and allow for useful complementary information sharing across experts.
The results establish that MosaicIMU provides a scalable pretraining-to-deployment paradigm for generalizable and adaptive neural inertial odometry. The experiments on four carrier categories show that our method achieves the best average generalization performance, reducing average ATE and RTE10s by 40% and 34%, respectively. p. 7 The results suggest that sample-adaptive feature-level composition provides a more flexible inertial prior than either carrier-specific training or fixed-head multi-carrier training. p. 5 p. 5 This framework is particularly effective on vehicle and pedestrian sequences, where heterogeneous motion patterns are difficult to capture with a fixed output partition. The gains are especially clear on vehicle and pedestrian sequences, where heterogeneous motion patterns are difficult to capture with a fixed output partition. The results establish that MosaicIMU provides a scalable pretraining-to-deployment paradigm for generalizable and adaptive neural inertial odometry.
The experiments on four carrier categories show that our method achieves the best average generalization performance, reducing average ATE and RTE10s by 40% and 34%, respectively. The results suggest that sample-adaptive feature-level composition provides a more flexible inertial prior than either carrier-specific training or fixed-head multi-carrier training. This framework is particularly effective on vehicle and pedestrian sequences, where heterogeneous motion patterns are difficult to capture with a fixed output partition. The gains are especially clear on vehicle and pedestrian sequences, where heterogeneous motion patterns are difficult to capture with a fixed output partition. The results establish that MosaicIMU provides a scalable pretraining-to-deployment paradigm for generalizable and adaptive neural inertial odometry.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive Feature Composition for Heterogeneous Carriers: The system can achieve sample-adaptive feature-level expert composition by using a prototype-based router to compose carrier-specific expert features, which allows for flexible inertial prior representation rather than fixed carrier heads.
-
Unseen Domain Adaptation via Lightweight Residual Branches: For unseen domains, the system can freeze the pretrained base model and introduce a new lightweight expert residual branch to learn new motion patterns efficiently while preserving the strong cross-carrier generalization learned during pretraining.
-
Resource-Constrained Online Adaptation: The system can implement router-guided sample selection to maintain a compact online buffer of informative samples for efficient incremental updates, ensuring performance on edge deployment with limited computational resources.
-
History-Aware State Estimation: The system integrates local velocity constraints with a history-aware Extended Kalman Filter (EKF) that uses segment-level average-velocity observations between cloned poses to enforce velocity consistency without introducing absolute global yaw observations.
-
Robust Velocity Prediction Frames: The system can utilize the local gravity aligned frame L0 for dense velocity predictions, which
avoids adding artificial global-yaw information to the EKF update
and provides afixed and gravity consistent reference for each window.
Sources
- KISS-IMU: Self-supervised Inertial Odometry with Motion-balanced Learning and Uncertainty-aware Inference
- i2Nav-Robot: A Large-Scale Indoor-Outdoor Robot Dataset for Multi-Sensor Fusion Navigation
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation