Research released between 16 and 22 September 2026 converged on a practical problem: embodied systems cannot assume that their observations arrive clean, complete, or in a fixed format. Three new arXiv preprints respond by stabilising irregular motion markers, inferring touch from ordinary video, and asking a robot to gather tactile evidence only when vision is insufficient.

DirtyMoCap reconstructs movement from unordered markers

On 17 September, Long Wang and colleagues released DirtyMoCap, a marker-layout-free motion-capture framework. Conventional optical motion capture expects a known arrangement of markers and relatively clean trajectories. DirtyMoCap instead accepts sparse, noisy, unordered point clouds whose layout can vary.

The system maps observations to a fixed collection of proxy anchors—skeletal joints and body-surface points—then fits an SMPL-H body model with a differentiable Gauss–Newton solver. The authors report better joint and vertex reconstruction than configuration-specific baselines across varied noisy layouts, plus as much as a 100-fold solver speed-up over their standard PyTorch implementation. They also apply the method to heterogeneous recordings of traditional Chinese martial arts.

The contribution is not a claim that marker cleanup has disappeared. It is a stronger intermediate representation: unreliable observations are converted into stable anatomical anchors before full-body reconstruction. This extends, rather than replaces, the issues discussed in Somatic-AI Lab’s account of why motion-capture suits remain difficult in practice.

TouchSight predicts contact without a glove at capture time

Also on 17 September, Danyan Zhou and colleagues released TouchSight, a monocular egocentric-video system for dense, full-hand contact-force prediction. The training evidence includes 500 hours of pressure-glove recordings. To bridge the visual difference between a gloved training hand and a bare hand, the team created TwinTouch-20H: 20 hours of paired video in which generative models re-render gloved recordings as bare-hand scenes while retaining the measured tactile labels.

The authors report improvement over earlier contact-prediction methods on OakInk2 and qualitative transfer to unseen natural video. That is promising, but it does not make inferred force equivalent to direct measurement. Generated appearance can expand the usable visual domain; the pressure glove remains the source of the training labels. Readers can compare that distinction with last week’s report on wrist pressure, hand pose, and contact force.

AVT-Fabric lets uncertainty decide when the robot should touch

On 18 September, Chang Gao and colleagues released AVT-Fabric, an active vision-and-touch system for comparing fabrics. It begins with RGB images, monitors answer confidence and logit separation, and requests another force-tagged GelSight observation only when the comparison remains uncertain.

On 400 held-out comparisons, the paper reports 98.0% accuracy while using an average of 1.60 of five available evidence stages. Its deployed robot achieved 78.1% pairwise-ranking accuracy and selected the correct fabric in seven of eight application scenarios. Those numbers describe the authors’ benchmark and deployment, not a general solution to material perception.

What connects these releases

All three projects treat sensing as a decision problem. DirtyMoCap asks which anatomical structure can survive corrupted markers. TouchSight separates directly measured training labels from inferred contact at capture time. AVT-Fabric asks whether another physical observation is worth its cost. For practitioners, the shared lesson is simple: record not only the model’s answer, but also which evidence was observed, inferred, generated, or actively requested.

References

  • Wang, L., Zhao, S., Yan, S., Yu, S., Li, X., Cai, Z., Hou, Y., & Xiu, Y. (2026). DirtyMoCap: Robust motion capture from unconstrained markers. arXiv. https://arxiv.org/abs/2609.19927
  • Zhou, D., Lu, J., Lin, J., Chen, T., Lyu, C., & Ding, W. (2026). TouchSight: Bare-handed tactile prediction from egocentric video via generative visual augmentation. arXiv. https://arxiv.org/abs/2609.20414
  • Gao, C., Chen, Z., Xia, S., Zhu, J., Deng, J., & Luo, S. (2026). AVT-Fabric: Active visuo-tactile perception via adaptive evidence selection for efficient robotic fabric comparison. arXiv. https://arxiv.org/abs/2609.21377