Group Inertial Poser

• Group Inertial Poser is a multi-person motion capture method that estimates full-body 3D pose and global translation by integrating sparse IMU data with UWB distance constraints.
• It employs a two-stage optimization strategy, first using structured state-space models for individual pose estimation then aligning poses in a shared world frame with pairwise sensor distances.
• Empirical evaluations show significant improvements, including up to 72% reduction in distance error over inertial-only baselines on both synthetic and real datasets.

Searching arXiv for the primary paper and closely related inertial poser work. Group Inertial Poser (GIP) is a multi-person motion-capture method that estimates full-body $3\text{D}$ pose and global translation for several people from sparse wearable inertial measurement units (IMUs) augmented with ultra-wideband (UWB) ranging. Its central objective is to recover both articulated body motion and shared-world spatial relationships between interacting people, addressing a core limitation of IMU-only systems: inertial cues are self-referential, provide no direct spatial reference for others, and therefore compromise translation estimates and accurate relative positioning between individuals over time (Xue et al., 24 Oct 2025).

1. Problem setting and motivation

GIP targets multi-person motion capture in unconstrained real-world settings using only body-worn sensors. The problem is motivated by two complementary limitations in prior capture paradigms. Traditional camera-based motion capture depends on fixed camera setups, limited capture range, and suffers from occlusion in crowded scenes or when people move close together. Sparse-IMU systems remove these environmental constraints, but they do not provide an external spatial reference and therefore drift in global translation. In the multi-person case, this weakness becomes more consequential: estimating each person independently can yield plausible single-person poses while still misplacing people relative to one another in the shared world frame (Xue et al., 24 Oct 2025).

The paper frames this as a deficiency in observability. IMUs measure relative acceleration and angular velocity, but not direct inter-person geometry. A multi-person capture system therefore needs constraints that go beyond body-internal kinematics. GIP introduces such constraints through pairwise sensor distances measured by UWB, including both same-person and between-person ranges. These distances provide information in meters and directly constrain interpersonal spacing, which is essential for preserving interaction structure such as proximity, approach, and separation (Xue et al., 24 Oct 2025).

This design distinguishes GIP from single-person inertial pose reconstruction. A plausible implication is that GIP is not merely extending an existing inertial poser with a second instance of the same model; it is reformulating motion capture so that inter-person geometry becomes an explicit input to estimation and optimization.

2. Sensor configuration and mathematical representation

For each user $i \in \{1,2\}$ and frame $t$, GIP uses sparse wearable sensors placed on the head, pelvis, wrists, and knees, giving $S=6$ sensing locations. The input variables are:

• orientations $R^i_t \in \mathbb{R}^{S\times 3}$,
• accelerations $A^i_t \in \mathbb{R}^{S\times 3}$,
• same-person sensor distances $D^i_t \in \mathbb{R}^{S\times S}$,
• between-person sensor distances $D^{12}_t \in \mathbb{R}^{S\times S}$.

The method predicts, for each person $i$,

• SMPL pose parameters $\mathbf{\Theta}^i \in \mathbb{R}^{N \times 3J}$,
• global translation $i \in \{1,2\}$0,

with $i \in \{1,2\}$1 joints and $i \in \{1,2\}$2 the sequence length. The model input is written as

$i \in \{1,2\}$3

This representation couples body-internal inertial measurements with geometric distance constraints across the full set of worn devices (Xue et al., 24 Oct 2025).

The use of UWB is central. Same-person distances help stabilize body configuration, while between-person distances provide direct cross-person spatial information. The paper emphasizes that UWB remains noisy, especially under non-line-of-sight conditions and body occlusion, but still contributes information that IMUs fundamentally lack: explicit spatial relationships in the world. That distinction is foundational for GIP’s subsequent optimization strategy (Xue et al., 24 Oct 2025).

3. Structured state-space estimation of individual motion

The first stage of GIP estimates each person separately: $i \in \{1,2\}$4 This stage uses a structured state-space model (SSM) inspired by S4. The continuous-time formulation is

$i \in \{1,2\}$5

and after zero-order-hold discretization,

$i \in \{1,2\}$6

The stated motivation is that SSMs can model long sequences efficiently and capture temporal dependencies better than recurrent models such as LSTMs in this setting (Xue et al., 24 Oct 2025).

The estimator combines multiple specialized modules rather than a single monolithic sequence model.

Component	Role
SSM-J	Estimates sensor positions from orientation and acceleration
GCN	Processes orientation and same-person distances to predict sensor positions
Adaptive fusion	Combines SSM-J and GCN outputs
SSM-R	Predicts joint angles
SSM-V	Predicts joint velocities
SSM-C	Predicts foot contact states
Physics optimizer	Enforces plausible motion

This composition indicates that GIP treats pose recovery as a fusion problem across temporal dynamics, graph-structured geometric relations, and physical plausibility. The output of this stage is full-body SMPL pose $i \in \{1,2\}$7 and translation $i \in \{1,2\}$8, but initially only relative to each person’s own frame (Xue et al., 24 Oct 2025).

The reliance on same-person UWB distances already differentiates the first stage from inertial-only baselines. A plausible implication is that same-person ranging is used not only as an auxiliary cue for global placement, but also as an internal regularizer on body configuration before any cross-person alignment is attempted.

4. Shared-world alignment and two-step translation optimization

After individual pose estimation, GIP reconstructs each person in a common world frame through a two-step optimization that uses between-person UWB distances. Predicted sensor positions are obtained from the estimated SMPL pose and translation: $i \in \{1,2\}$9 where $t$0 is the person’s initial world position and $t$1 is the SMPL forward-kinematics mapping from pose to sensor locations. Predicted between-person distances are then

$t$2

and the paper summarizes the mapping as

$t$3

This explicitly converts articulated pose and translation hypotheses into the measurable domain of UWB pairwise distances (Xue et al., 24 Oct 2025).

The first optimization stage solves the unknown initial relative placement. One person is fixed at the origin,

$t$4

and the second person’s initial relative position $t$5 is optimized by minimizing the mismatch between predicted and measured between-person distances: $t$6

The second stage jointly refines the full trajectories: $t$7 with

$t$8

The first term enforces agreement with UWB distances, while the regularizers constrain velocity and acceleration mismatch. The paper states that directly optimizing the whole trajectory without the initial-position step is unstable and can converge to unrealistic paths; the two-stage design is therefore presented as a conditioning strategy rather than a mere implementation convenience (Xue et al., 24 Oct 2025).

5. GIP-DB, evaluation protocol, and empirical findings

GIP is evaluated on both synthetic and real data. Training is performed from scratch on AMASS, with synthetic IMU and UWB signals generated from AMASS and InterHuman by simulation. Evaluation uses InterHuman for synthetic two-person interactions and GIP-DB for real-world two-person IMU+UWB motion tracking (Xue et al., 24 Oct 2025).

GIP-DB is introduced as the first IMU+UWB dataset for two-person tracking. It contains 7 pairs of participants, 14 participants total, with 10 male and 4 female participants and heights ranging from 160 cm to 195 cm. The dataset comprises over 200 minutes of motion and includes walking, stretching, jogging in place, close conversation, sparring, handshakes, and dancing. Each participant wore an Xsens MVN Awinda suit with 17 IMUs for ground-truth SMPL pose and 6 custom sensors on the head, pelvis, wrists, and knees. Each custom sensor included a DWM3000 UWB radio and an LSM6DSL 6DoF IMU. UWB ranging was recorded at 40 Hz, IMU data at 104 Hz, and ground-truth translation came from a 20-camera OptiTrack system. Sequences began and ended with a T-pose and a jump for synchronization. The average UWB RMSE is reported as about 5 cm for same-person measurements and about 15 cm for between-person measurements, highlighting the challenge of real-world non-line-of-sight ranging (Xue et al., 24 Oct 2025).

The evaluation compares GIP against PIP and UIP. For fairness, PIP and UIP are given ground-truth initial translation, since they cannot predict it natively, and PIP is augmented with sensor-distance information like UIP. Metrics include SIP Error, Angle Error, Joint Error, Trans Error @ 3m / 6m, and Dist Err @ 4s, 8s, 12s, 16s, 20s. The distance-error metric is particularly important because it directly measures whether interpersonal spacing is preserved over time (Xue et al., 24 Oct 2025).

On InterHuman, GIP outperforms PIP and UIP on all pose and translation metrics, with key reported improvements of 22% lower full-body joint angle error versus UIP and 33% lower joint angle error versus PIP. The paper further states that GIP’s distance errors are around only a few centimeters, while PIP and UIP are often tens of centimeters off and can exceed 60–80 cm over longer horizons. On the real-world GIP-DB benchmark, GIP again shows better pose accuracy and lower translation error, including a 72% reduction in distance error at 20s compared to baselines. Finetuning on GIP-DB further improves angular predictions. The initialization study reports that the initial-position optimizer yields results close to those using ground-truth initialization (Xue et al., 24 Oct 2025).

Ablation results attribute these gains to all major components. Removing initial position optimization substantially degrades translation performance; removing trajectory optimization worsens stability; and replacing the SSM with an LSTM increases angle and translation errors. The method is also shown to extend beyond two people: in a synthetic four-person setting, translation error decreases from $t$9 to $S=6$0 to $S=6$1 to $S=6$2 m as the number of people increases from 1 to 4, suggesting that additional cross-person distance constraints improve global trajectory estimation (Xue et al., 24 Oct 2025).

6. Position within the inertial poser literature, limitations, and significance

GIP belongs to the broader inertial poser family but addresses a different target from earlier systems. Transformer Inertial Poser (TIP) reconstructs full-body motion from six sparse IMUs while simultaneously generating plausible terrain, using a conditional Transformer decoder, stationary body points, and terrain-aware correction; its emphasis is temporal consistency, drift correction, and non-flat terrain for a single subject (Jiang et al., 2022). Physical Inertial Poser (PIP) combines a causal RNN-based kinematics estimator with a physics-aware optimizer to estimate pose, translation, joint torques, and ground reaction forces from six IMUs, but it remains a single-subject system and assumes flat ground (Yi et al., 2022). Against that background, GIP’s distinctive contribution is not a new single-person kinematic prior, but the introduction of UWB-based inter-sensor distance constraints and two-step shared-world translation optimization for multiple people.

This distinction also clarifies a common source of confusion in the inertial-poser nomenclature. GIP is specifically the method titled "Group Inertial Poser" (Xue et al., 24 Oct 2025); it is separate from "Transformer Inertial Poser" (Jiang et al., 2022), "Physical Inertial Poser" (Yi et al., 2022), "Progressive Inertial Poser" (Zhu et al., 8 May 2025), and other later variants that focus on minimal sensor count, garment-based sensing, shape awareness, or pressure-based dynamics rather than multi-person shared-world reconstruction.

The paper explicitly identifies several limitations. UWB noise remains a challenge, especially under non-line-of-sight and body-occluded interactions. The model assumes a mean body shape and does not estimate per-person shape variation. The optimization introduces computational overhead, although it is reported to require fewer than 10 iterations and about 2.04 seconds for a 30-second sequence. The method also does not explicitly address foot sliding (Xue et al., 24 Oct 2025).

Within those constraints, GIP points toward practical multi-person motion capture in crowded indoor spaces, social VR, sports or dance interactions, rehabilitation and movement analysis, collaborative robotics, wearable interaction research, and privacy-sensitive environments where cameras are undesirable. A plausible implication is that its most consequential advance is not simply improved trajectory accuracy, but the recovery of meaningful interpersonal dynamics in a common coordinate frame, which earlier single-person inertial posers were not designed to model.