IMUDINO: The First General-Purpose Inertial Sensing Foundation Model, Breaking Free from Sensor Placement and Sampling Rate Constraints

Inertial sensors (IMUs) are found in nearly every smartphone, every pair of earbuds, and every robot, yet compared to foundation models in vision and natural language, research on foundation models for inertial sensing has long remained fragmented. A recent paper proposes IMUDINO, which aims to change this—it claims to be the first general-purpose foundation model that achieves competitive performance across a variety of inertial sensing tasks, covering human activity recognition (HAR), health monitoring, motion capture, and odometry.

Why Inertial Sensing Needs Its Own Foundation Model

Past pre-trained models for inertial data have tended to focus on only one subdomain—for example, only human activity recognition, or only gait analysis. The direct consequence of this narrow scope is that on tasks like HAR, these models sometimes perform even worse than carefully tuned supervised learning methods. Compared with the cross-task generalization demonstrated by vision and language foundation models, "foundation models" in inertial sensing are foundation models in name only.

Another overlooked pain point is at the engineering deployment level. Many existing models require fixed input lengths and sampling rates; once the device changes or the sampling configuration shifts, the model needs to be re-adapted. Even trickier is sensor placement generalization—when a model is transferred from the wrist to a pocket or ankle, performance often drops sharply. These limitations make it difficult for inertial foundation models to be truly "general."

Learning Motion Semantics with Self-Distillation

The core idea of IMUDINO is to transfer the DINO self-supervised learning paradigm to inertial data. DINO learns representations in a self-distillation manner in the vision domain without requiring manual annotation. Applying it directly to IMU data is not feasible, because inertial signals lack the spatial structure of images. The research team designed a set of "semantics-preserving motion transformations" that construct self-supervised signals by simulating variations in motion patterns and sensor configurations.

The key to this transformation strategy is that the transformed data remains semantically the same action or the same motion state, but the signal morphology changes. The model is forced to learn representations that are invariant to these changes, thereby gaining generalization ability across users and sensor configurations. This goes a step further than simple data augmentation, because it directly encodes the sources of variation that genuinely exist in inertial sensing.

The DINO self-supervised learning paradigm: DINO (self-distillation with no labels) was proposed by Meta AI in 2021, with its core being self-distillation between a student network and a teacher network—the teacher network is an exponential moving average (EMA) of the student network's weights, the two receive different augmented views of the same image, and the student is trained to match the probability distribution output by the teacher. To avoid model collapse (all inputs mapping to the same representation), DINO employs two techniques: centering and sharpening—centering prevents one dimension from dominating, while sharpening uses a low-temperature softmax to make the distribution more peaked. DINO can learn high-quality visual representations usable for classification, segmentation, and detection without needing negative sample pairs and large memory banks. What IMUDINO borrows is precisely this idea of "learning invariant representations from view consistency without manual annotation," except that image augmentations are replaced with semantics-preserving motion transformations.

Frequency-Aware Positional Encoding and Variable-Length Inputs

At the architectural level, IMUDINO adopts a plain Transformer encoder without introducing complex specialized modules. The real design highlight lies in frequency-aware positional encoding. This design allows the model to natively support varying input lengths and sampling rates, without the need to resample data to a uniform specification during preprocessing.

The paper particularly emphasizes that this simple setup "has proven remarkably effective," being able to learn both global and dense features and to generalize across users, sensor setups, and tasks. In robotic odometry tasks, this generalization ability even extends across different embodiments—that is, the same model can be used across different robot platforms. This is quite significant for practitioners in robotics, because the differences in IMU mounting positions and data characteristics across platforms are often substantial.

The relationship between positional encoding and variable-length inputs: Transformers are inherently insensitive to input order, and positional encoding is responsible for injecting sequence position information. Standard learnable positional encodings have a fixed length and cannot be used directly when the input length varies; although sinusoidal positional encodings can extrapolate, their frequency meaning becomes misaligned when the sampling rate changes. The essence of inertial signals is time series, whose physical meaning is determined by the sampling frequency—the same action sampled at 50Hz and 200Hz has different numbers of sample points, but the corresponding real time span may be the same. The idea of frequency-aware positional encoding is to make the positional encoding explicitly carry time or frequency information, so that the model can understand the real time scale behind "the nth sample point," thereby maintaining representation consistency when input length and sampling rate vary. This is also the key reason why IMUDINO requires no resampling during preprocessing and can natively adapt to different device configurations.

Multi-Task Coverage and the Potential of a General Backbone

In terms of task coverage, IMUDINO targets four directions simultaneously: HAR, health, motion capture, and odometry. This cross-subdomain design approach is consistent with the evolutionary path of vision foundation models: first have a sufficiently general backbone network, then perform lightweight adaptation for downstream tasks. The authors hope that IMUDINO can become the backbone model in the inertial sensing field and drive the development of general inertial foundation models.

It should be noted that the paper abstract does not disclose specific performance numbers or details of comparison baselines. From the wording, its "competitive" positioning holds simultaneously across multiple tasks rather than being a single-point breakthrough. For developers focused on edge computing and wearable devices, a model that can natively handle variable-length inputs, adapt to different sampling rates, and is insensitive to sensor placement may be more practically appealing in engineering terms than pure accuracy improvements.

Inertial odometry and cross-embodiment generalization: Inertial odometry refers to inferring a carrier's motion trajectory solely from IMU accelerometer and gyroscope data, without relying on GPS or vision. Its difficulty lies in the fact that integration errors accumulate and drift over time, and the motion patterns and IMU mounting methods differ greatly across carriers (humans, wheeled robots, quadruped robots). Cross-embodiment generalization means that the same model can handle the IMU data distributions of different platforms without retraining for each robot platform. This capability is particularly important for robotics: there are many types of real robot platforms, and collecting annotated data and training odometry models separately for each platform is extremely costly; a general backbone network can significantly lower the deployment barrier.

What It Means for Developers

Foundation models for inertial sensing have long been absent, partly because data annotation is costly and device heterogeneity is strong. IMUDINO bypasses the annotation bottleneck with a self-supervised approach and bypasses device heterogeneity with frequency-aware encoding. If these two design choices can be validated and supported by open source in subsequent work, they will lower the barrier for wearable and robotics applications to adopt foundation models. Currently, this work is at the paper publication stage; whether the code and model weights will be open-sourced remains unclear, and its practical usability remains to be seen.

One-line Summary

IMUDINO brings the DINO self-supervised paradigm to inertial sensing, creating the first multi-task general IMU foundation model.

Paragraph Summary

IMUDINO is the first general-purpose foundation model for inertial sensing, covering four types of tasks: human activity recognition, health monitoring, motion capture, and robotic odometry. To address the narrow scope and poor generalization of existing inertial pre-trained models, it transfers the DINO self-distillation paradigm from the vision domain to IMU data, designing semantics-preserving motion transformations to construct self-supervised signals and forcing the model to learn representations invariant to changes in users and sensor configurations. Architecturally, it adopts a plain Transformer encoder, with the core innovation being frequency-aware positional encoding, which enables the model to natively support variable-length inputs and different sampling rates without resampling. In robotic odometry, it also demonstrates generalization ability across embodiments. The paper abstract does not disclose specific performance numbers or open-source plans, and its practical usability remains to be seen.