Step2Motion: Locomotion from Wearable Insoles

• Step2Motion is a framework that reconstructs full-body locomotion using wearable insole sensors, combining pressure and IMU signals.
• It employs a diffusion-based pose model and a specialized Transformer for root displacement to enhance leg accuracy and temporal consistency.
• Empirical results on multiple datasets demonstrate its effectiveness in diverse motion scenarios, while noting limitations in reconstructing upper-body dynamics.

Searching arXiv for the primary Step2Motion paper and closely related work on motion reconstruction from wearable/pressure sensing. Step2Motion is a locomotion reconstruction method that infers full-body human motion from multi-modal pressure-sensing insoles. It is presented as the first approach to reconstruct human locomotion from multi-modal insole sensors, using pressure and inertial data—accelerations and angular rates—captured by the insoles to reconstruct human motion across diverse locomotion styles, including walking, jogging, moving sideways, on tiptoes, slightly crouching, and dancing. The method is motivated by the observation that foot–ground interaction encodes contact, push-off, balance, and locomotion direction, while insoles remain wearable, cheap, unobtrusive, and usable in outdoor or in-the-wild settings without line-of-sight constraints (Ponton et al., 26 Oct 2025).

1. Task formulation and reconstruction objective

Step2Motion formulates motion over a temporal window of length $W$ as a sequence of poses

$\mathbf{p}=(\mathbf{d}, \mathbf{j}),$

where $\mathbf{d}\in \mathbb{R}^3$ is the root displacement from the previous frame in world coordinates, and $\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$ are the relative 3D joint positions for the remaining $J-1$ joints. The global root trajectory is recovered by cumulative summation,

$\sum^i_{f=0}\mathbf{d}^{(f)}.$

The synchronized insole observation at frame $i$ is

$$\mathbf{c}^{(i)} = (\mathbf{c_p^L}, \mathbf{c_a^L}, \mathbf{c_r^L}, c_f^L, \mathbf{c_c^L}, \mathbf{c_p^R}, \mathbf{c_a^R}, \mathbf{c_r^R}, c_f^R, \mathbf{c_c^R}) \in \mathbb{R}^{50},$$

with 25 features per foot. The learning problem is therefore to synthesize the pose sequence $(\mathbf{p}^{(i)})_{i=1}^W$ from $(\mathbf{c}^{(i)})_{i=1}^W$ (Ponton et al., 26 Oct 2025).

A central difficulty is that the insole signal is ambiguous: similar measurements can correspond to multiple plausible full-body poses. The paper therefore treats motion reconstruction as a one-to-many problem rather than a purely deterministic regression problem. This motivates the use of a diffusion-based pose model for relative joint reconstruction and a separate predictor for root displacement, rather than a single monolithic regressor (Ponton et al., 26 Oct 2025).

2. Sensing setup and motion-relevant observables

The hardware platform is based on Moticon OpenGo Sensor Insoles. Each insole contains 16 pressure sensors and one IMU. The IMU provides linear acceleration $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$0 and angular rate $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$1; the method also uses total force $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$2 and center of pressure $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$3 in normalized coordinates $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$4. Across both feet, these measurements form the 50D per-frame signal used for pose reconstruction (Ponton et al., 26 Oct 2025).

The modality design is task-specific. For pose reconstruction, the full 50D vector is used because pressure, IMU, force, and center-of-pressure encode complementary information about support, local balance, and stance. For root displacement prediction, the paper often uses only the IMU data $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$5, because pressure can cause overfitting in that subtask. The acceleration is transformed from local to world coordinates by integrating angular rate, although some controlled experiments use ground-truth rotation to isolate model performance from IMU integration errors (Ponton et al., 26 Oct 2025).

This sensing formulation makes the feet the dominant observation interface. A plausible implication is that Step2Motion is strongest when lower-body dynamics and contact transitions dominate the motion signature, and weaker when large components of the target motion are kinematically distant from the feet. The paper’s later limitation analysis is consistent with that interpretation (Ponton et al., 26 Oct 2025).

3. Architecture: diffusion pose reconstruction with a separate displacement branch

Step2Motion has two main components. The first is a diffusion-based pose reconstruction model that reconstructs the pose sequence $\mathbf{p}=(\mathbf{d}, \mathbf{j}),$6 from a noisy latent sequence. Its forward process is DDPM-style,

$\mathbf{p}=(\mathbf{d}, \mathbf{j}),$7

$\mathbf{p}=(\mathbf{d}, \mathbf{j}),$8

and the reverse process is

$\mathbf{p}=(\mathbf{d}, \mathbf{j}),$9

approximated by a neural network

$\mathbf{d}\in \mathbb{R}^3$0

which predicts the denoised pose directly at each step rather than predicting noise (Ponton et al., 26 Oct 2025).

The second component is a separate Transformer-based displacement predictor that regresses root displacements $\mathbf{d}\in \mathbb{R}^3$1 from IMU data. This separation is architecturally important: the paper reports that integrating displacement prediction into the diffusion pathway is worse than using a dedicated displacement network, which suggests that root motion is sufficiently structured and sensor-correlated to benefit from a specialized regression branch (Ponton et al., 26 Oct 2025).

The pose model uses body-part partitioning. The motion sequence is split into

$\mathbf{d}\in \mathbb{R}^3$2

corresponding to left leg, right leg, and remaining body. These parts are projected to an embedding space of dimension $\mathbf{d}\in \mathbb{R}^3$3, augmented with sinusoidal positional encodings, and concatenated into

$\mathbf{d}\in \mathbb{R}^3$4

The Transformer then performs temporal sequence modeling with self-attention, diffusion timestep embeddings, and feed-forward layers (Ponton et al., 26 Oct 2025).

Conditioning from the insoles is also decomposed. The sensor stream is partitioned into 8 components total—4 per foot—based on toe pressure, heel pressure, IMU acceleration plus angular rate, and force plus center of pressure. Each component is embedded by a small MLP into $\mathbf{d}\in \mathbb{R}^3$5, and a custom multi-head cross-attention uses them as separate heads: $\mathbf{d}\in \mathbb{R}^3$6

$\mathbf{d}\in \mathbb{R}^3$7

The rationale is explicitly modality-aware: different sensor streams matter for different actions, with pressure often more informative for stationary balance or squat-like motions, and IMU more informative for translation direction and dynamic locomotion (Ponton et al., 26 Oct 2025).

For long motion sequences, the method adopts an autoregressive or inpainting-inspired strategy to generate temporally coherent motions beyond a single window (Ponton et al., 26 Oct 2025).

4. Losses, datasets, and evaluation protocol

The pose diffusion model is trained with mean absolute error between predicted and ground-truth pose sequences. The displacement predictor uses a composite loss consisting of per-frame MSE plus a cumulative-sum regularizer,

$\mathbf{d}\in \mathbb{R}^3$8

with $\mathbf{d}\in \mathbb{R}^3$9. This second term directly penalizes accumulated trajectory drift, so optimization targets not only local displacement accuracy but also long-horizon trajectory consistency (Ponton et al., 26 Oct 2025).

Experiments use two datasets. UnderPressure is a public dataset with 9 subjects and about 5 hours of mocap plus insole data, with limited motion variety concentrated on walking, jogging, and jumping. The Step2Motion dataset is newly recorded by the authors and contains 8 subjects, about 3.6 hours total, and approximately 25-minute continuous recordings per subject. Its motion inventory is broader, including walking, jogging, jumping, squatting or crouching, sideways movement, tiptoeing, dancing, and in-the-wild locomotion sequences. Train and test subjects are separated, and test motions are held out from training. The sampling rate is 30 Hz and the temporal window length is 100 frames (Ponton et al., 26 Oct 2025).

The evaluation metrics are split according to subtask. Pose reconstruction uses MPJPE, MPJPE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$0, and MPJVE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$1; displacement evaluation uses MRPE. For pose accuracy analysis on UnderPressure, acceleration is transformed to world space using ground-truth rotational data to remove IMU integration noise. Pose evaluation is root-aligned to ground truth, whereas displacement evaluation compares root trajectory directly (Ponton et al., 26 Oct 2025).

5. Quantitative behavior, ablations, and empirical scope

For pose reconstruction, the paper compares against an MLP baseline, a Transformer baseline, a version without the proposed insole multi-head attention, and the full model. On UnderPressure, the full model achieves MPJPE 7.2, MPJPE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$2 6.5, and MPJVE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$3 26.1. On the Step2Motion dataset, it achieves MPJPE 11.4, MPJPE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$4 12.3, and MPJVE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$5 50.4. The reported pattern is that the Transformer baseline is worst, especially on velocity, the MLP often oversmooths motion, and the full model is best overall, particularly for leg accuracy and temporal consistency (Ponton et al., 26 Oct 2025).

For displacement prediction, the paper compares double integration, MLP, a combined variant that predicts displacement inside diffusion, only-pressure, a version without cumulative-sum loss, pressure plus IMU, and only-IMU variants. The main qualitative findings are that double integration is very poor, especially on the more diverse Step2Motion dataset; removing the cumulative-sum term causes drift; only IMU is best for root motion on the more diverse dataset; pressure helps in some stationary cases but pressure-only is insufficient for general root displacement; and integrating displacement into the diffusion branch is worse than using a dedicated predictor (Ponton et al., 26 Oct 2025).

The long-horizon behavior is illustrated by an in-the-wild 60 m jogging loop. The full method yields a final drift of 0.75 m, approximately 1.25%, compared with 9.8 m for double integration, 3.85 m for MLP, 4.6 m for the combined variant, 2.45 m without the cumulative-sum term, and 3.25 m for pressure plus IMU; the only-pressure setting produces no movement (Ponton et al., 26 Oct 2025).

A separate dance evaluation underscores the need for both modalities and the specialized fusion mechanism. On a small dance dataset, the version without insole multi-head attention records MPJPE 31.6, MPJPE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$6 34.4, and MPJVE$\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$7 307.1; only-pressure gives 25.7, 26.1, and 296.9; only-IMU gives 21.3, 22.7, and 320.9; and the full model achieves 8.2, 6.9, and 34.2. The paper interprets this as evidence that both modalities plus the specialized fusion design are necessary for hard motions such as dancing (Ponton et al., 26 Oct 2025).

These ablations clarify the division of labor among the sensor streams. Walking and jogging rely strongly on IMU for direction and translation; squatting, crouching, and stationary balance rely more on pressure because posture is encoded in load distribution; tiptoeing and complex balance depend on pressure-sensitive foot-contact patterns; and dancing requires both modalities. This suggests that Step2Motion is not simply an insole-conditioned sequence model, but a modality-adaptive reconstruction framework whose gains depend on action-dependent sensor relevance (Ponton et al., 26 Oct 2025).

6. Limitations and position within adjacent motion research

The paper explicitly identifies several limitations. IMU drift remains a major issue, especially for root motion and global orientation. Body parts far from the feet, such as head and arms, are weakly constrained by insole data, so upper-body reconstruction is less reliable. Some motions, especially squats and similar actions, are harder to recover accurately when IMU signals are noisy or drifted. The pressure split is still coarse, using only toe-versus-heel partitioning. The displacement predictor can overfit if pressure is included too aggressively, particularly with limited training data. The paper therefore points to more data, synthetic insole generation, additional sensors such as egocentric cameras or more IMUs, larger motion priors, and finer-grained pressure-region modeling as future directions (Ponton et al., 26 Oct 2025).

Within adjacent literature, Step2Motion occupies a specific niche: full-body locomotion reconstruction from wearable insole sensing. A closely related but distinct line is pressure-conditioned motion synthesis from floor sensors. “Pressure2Motion” formulates a pressure-plus-text problem using ground pressure sequences rather than wearable insoles, and introduces a hierarchical diffusion model with a dual-level pressure feature extractor and the MPL benchmark (Li et al., 7 Nov 2025). This suggests that Step2Motion and Pressure2Motion address neighboring sensing regimes—wearable foot–ground interaction versus environmental pressure capture—while sharing the broader premise that contact signals are rich motion descriptors.

The name can also be confused with other “stepwise” motion formulations that are technically unrelated. In multiple object tracking, TrackSSM introduces a Step-by-Step Linear ($\mathbf{j}\in \mathbb{R}^{(J-1)\times 3}$8) training strategy that decomposes one box transition into a chain of smaller regression steps for trajectory prediction (Hu et al., 2024). In human motion generation, “Progressive Human Motion Generation Based on Text and Few Motion Frames” proposes a multi-stage diffusion process that fills low-uncertainty frames before high-uncertainty ones in a Text-Frame-to-Motion setting (Zeng et al., 17 Mar 2025). These works indicate that stepwise supervision and staged generation are recurring ideas across motion modeling, but Step2Motion specifically denotes locomotion reconstruction from multi-modal pressure-sensing insoles rather than generic stepwise motion transfer or progressive generation.

Taken together, the evidence positions Step2Motion as a sensor-driven motion reconstruction framework whose principal contribution is methodological rather than purely architectural: it shows that pressure insoles with embedded IMUs can support plausible full-body locomotion reconstruction when ambiguity is handled with diffusion modeling, multimodal fusion is made body-part- and sensor-aware, and root motion is delegated to a dedicated trajectory predictor (Ponton et al., 26 Oct 2025).