Ultra Diffusion Poser: Diffusion-Based Human Motion Tracking From Sparse Inertial Sensors and Ranging-Based Between-Sensor Distances

Abstract: Methods using inertial measurement units (IMUs) provide a wearable alternative to camera-based motion capture. To mitigate drift from inertial signals, recent sparse inertial pose estimators integrate inter-sensor distances measured by ultra-wideband (UWB) ranging. So far, UWB distances have only been used as an additional input feature, ignoring the physical constraints they impose on sensor positions. However, these distances can also be used to reconstruct the underlying 3D sensor layout, which in turn provides more informative input for pose reconstruction. We propose Ultra Diffusion Poser, a diffusion model that explicitly models these geometric constraints. It includes a Spatial Layout Module that analytically reconstructs the 3D sensor positions from UWB measurements. These sensor positions are used alongside IMU signals and UWB distances as a conditioning signal during diffusion. Still, network predictions can violate inter-sensor distance measurements. To address this, we introduce UWB-Diffusion Guidance, which encourages alignment between predicted poses and measured distances during diffusion sampling. Together, these contributions enable our model to achieve state-of-the-art performance, reducing joint position error by up to 22% over prior work.

• The paper introduces a diffusion-based motion tracker that reconstructs 3D sensor layouts from UWB distances and uses sampling-time guidance to enforce geometric consistency, reducing joint position error by up to 22%.
• UDP achieves 3.42 cm joint position error on DIP-IMU, improves results across multiple benchmarks, and runs at 0.151 ms per frame with five-step DDIM sampling—about 50 times faster than UIP.
• The method delivers smoother and more accurate poses but becomes sensitive to noisy UWB measurements, requires guidance-strength tuning, and does not guarantee physical constraints such as contact stability or zero foot skating.

Overview and motivation

Sparse inertial motion capture with six body-worn IMUs suffers from drift because inertial signals provide no absolute inter-frame observations. Recent systems fuse ultra-wideband (UWB) ranging, which yields non-drifting pairwise distances between sensors, but prior work (UIP, UMotion, GIP) treats these distances merely as auxiliary input features. This has two consequences: the physical constraints the distances impose on sensor positions are never exploited to recover the underlying 3D sensor layout, and predicted poses can directly contradict measurements—for example, estimating wrist positions 70 cm apart when UWB reports 90 cm. Ultra Diffusion Poser (UDP) addresses both shortcomings within a diffusion-based pose estimator (2606.02153).

Method

UDP is a fully learnable autoregressive diffusion-inpainting model that maps global sensor orientations $\mathbf{R}$, accelerations $\mathbf{A}$, and pairwise distance matrices $\mathbf{D}$ from $k=6$ sensors to SMPL parameters $(\Theta_t, \mathbf{T}_t, \beta)$. Two components explicitly model the geometric content of UWB:

Spatial Layout Module (SPL): The SPL applies classical metric multidimensional scaling (MDS) to the distance matrix, recovering 3D sensor positions in closed form via eigendecomposition of the double-centered squared-distance Gram matrix. Because MDS output is defined only up to rotation and reflection, a learnable Rotation Estimator (a 3-layer LSTM trained end-to-end with the diffusion model) orients both the normalized layout and its mirrored candidate, and predicts per-sensor residual corrections to compensate for MDS noise. The resulting oriented layouts $(P_{SPL}, \bar{P}_{SPL})$ serve as a strong geometric prior alongside raw IMU and UWB signals in the conditioning tensor.

UWB-Diffusion Guidance: Even with strong conditioning, sampled poses can violate measured distances. Following classifier guidance, UDP shifts the sampling mean by the gradient of a guidance loss $$\epsilon_{uwb} = \sum_{i<j} \|\hat{d}_{ij} - d_{ij}\|^2$$, where predicted distances are computed via forward kinematics on the denoised pose estimate inside the sampling loop. Gradients for root translation are zeroed since UWB constrains only inter-body geometry. Unlike frame-wise post-optimization—which the authors argue produces jitter and infeasible joint angles under sensor noise—guidance acts throughout sampling and preserves smoothness.

The denoiser is an LSTM-based DDPM with diffusion inpainting over a history of $N_H=30$ frames for temporal consistency, trained without the expensive forward-kinematics loss used by DiffusionPoser; instead it uses direct SMPL angle, translation, and velocity losses. Notably, the authors report that adding an FK loss actually degrades accuracy marginally while tripling training time, attributing this to unstable gradients through the kinematic chain (up to 10× higher gradient norms than direct supervision). The full model is compact at 11.3M parameters and trains in roughly five hours on an RTX 4090.

Results

UDP achieves state-of-the-art joint position error across nearly all evaluation settings, reducing JPE by up to 22% over prior work and by up to 35% relative to IMU-only methods. On DIP-IMU it reaches 3.42 cm JPE and 10.39° SIP error, versus 4.38 cm / 13.20° for UIP, while also producing the lowest jitter (0.125 m/s³ against ground truth of 1.830). On DanceDB and TotalCapture it attains the lowest SIP, GAE, and JPE simultaneously. On real-world data, UDP is state-of-the-art on GIP-DB in both zero-shot (8.86 cm JPE) and fine-tuned settings (6.68 cm JPE, 15.33° SIP), and on UIP-DB after fine-tuning (9.04 cm JPE versus UIP's 10.65 cm).

A noise-blending study explains the one exception: on UIP-DB zero-shot, where UWB errors exceed 17 cm, UIP slightly outperforms UDP (-1.07 cm JPE), but as measurement noise is reduced UDP's advantage grows monotonically to +1.40 cm JPE and +4.59° GAE with perfect distances. This confirms that UDP's tight architectural coupling to UWB makes it more sensitive to noisy ranging—a trade-off the authors state plainly—and implies its results track the quality of modern UWB hardware.

Ablations on TotalCapture show complementary contributions: removing the SPL raises SIP by 7%, removing guidance raises SIP by 5% and JPE by 7%, and removing both increases JPE by 17%. Guidance strength must be tuned to expected noise ($\lambda=500$ for clean synthetic conditions, $\lambda=50$ for real-world datasets).

Efficiency and design ablations

Because UDP avoids UIP's serial CPU-bound QP physics optimizer, inference is substantially faster: with 5-step DDIM sampling, UDP runs at 0.151 ms per frame—roughly 50× faster than UIP—while SIP error rises only 4.8% relative to 50-step DDPM. Additional ablations support several design choices:

• LSTM over transformer: replacing the LSTM denoiser with a transformer encoder increases JPE by 27%, suggesting local neighborhood priors suit this dataset scale.
• Square-root noise schedule: counter to the hypothesis that gentler schedules would protect kinematic-chain structure, the aggressive square-root schedule outperforms cosine, exponential, and linear schedules.
• Absolute translation representation: predicting window-relative absolute positions rather than per-frame velocities reduces translation drift by 16 cm over 5 m.
• No lookahead needed: evaluating only the last frame of the window (fully online) costs about 2.7% SIP, and UDP still outperforms baselines in that setting.
• Body shape conditioning: providing ground-truth $\mathbf{A}$0 reduces SIP error by 10%.

On translation specifically, results are mixed: UDP achieves lower cumulative translation error than all baselines on DanceDB, where improved foot placement helps gait estimation, but physics-based methods (PIP, PNP, GlobalPose) retain an advantage of roughly 5 cm over 5 m on TotalCapture due to explicit contact constraints.

Limitations and open questions

The paper concedes several limitations. First, UDP's reliance on UWB makes it fragile under high ranging noise; the guidance coefficient must be manually adjusted to the expected noise level, and principled modeling of sensor uncertainty remains open. Second, being purely data-driven, UDP enforces no physical constraints—it cannot guarantee zero foot skating, ground-penetration avoidance, or strict dynamic consistency, so applications prioritizing physical plausibility may still favor physics optimizers; whether physics-based guidance terms can be integrated into the diffusion framework is left unexplored. Third, like prior work, UDP is restricted to a fixed six-sensor configuration, and extension to arbitrary sensor placements is unresolved. Finally, the approach addresses only intra-body constraints; extending SPL and guidance to inter-person distances (as in GIP) or environment-fixed UWB beacons is proposed but not evaluated.

Conclusion

UDP demonstrates that treating UWB inter-sensor distances as geometric constraints—both through closed-form 3D layout reconstruction conditioned into a diffusion model and through FK-based guidance during sampling—yields measurable gains over using them as auxiliary features, with up to 22% lower joint position error, smoother motion, and order-of-magnitude faster inference than optimization-based alternatives. Its sensitivity to noisy ranging and lack of physical plausibility guarantees define the boundaries of these gains and the questions the paper leaves open.