The arXiv agent has returned with a substantially larger set of confirmed papers. Here is the addendum — these items supplement the report above. Together they form the complete week's digest.
arXiv Addendum — Additional Confirmed Papers (2026-03-31 → 2026-04-06)
The first scan returned eight papers; the full arXiv pass confirmed seventeen more within the window. The additions below are ordered by field relevance.
[8] Nonlinear Methods for Analyzing Pose in Behavioral Research Carter Sale, Margaret C. Macpherson, Gaurav Patil, Kelly Miles, Rachel W. Kallen, Sebastian Wallot, Michael J. Richardson Source: arXiv cs.HC · Date: 1 April 2026 URL: https://arxiv.org/abs/2604.01453
A general-purpose pipeline for human pose data combining preprocessing, dimensionality reduction, and recurrence-based time-series analysis (recurrence quantification analysis, fractal scaling) to characterise movement dynamics. Three case studies span facial movement, full-body motion, 2D/3D data, and individual and group interactions. Directly relevant to somatic AI: it applies dynamical systems mathematics to body data, enabling analysis of movement quality, coordination, and spontaneous self-organisation—phenomena central to somatic practice that metric-based approaches miss.
Sale, C. et al. (1 April 2026). Nonlinear Methods for Analyzing Pose in Behavioral Research. arXiv:2604.01453. https://arxiv.org/abs/2604.01453
[9] Sparkle: A Robust Representation for Point Cloud-Based Human Motion Capture Yiming Ren, Yujing Sun, Aoru Xue, Kwok-Yan Lam, Yuexin Ma Source: arXiv cs.CV · Date: 1 April 2026 URL: https://arxiv.org/abs/2604.00857
Combines skeletal joints and surface anchors with kinematic-geometric separation to enable high-fidelity, robust 3D human motion capture from LiDAR/point-cloud data—maintaining accuracy under domain shifts, noise, and occlusion across sensor types. Significant for somatic AI as it opens high-quality motion capture to portable, low-cost sensors in uncontrolled environments: studios, outdoor settings, live performance spaces.
Ren, Y. et al. (1 April 2026). Sparkle. arXiv:2604.00857. https://arxiv.org/abs/2604.00857
[10] EgoSim: Egocentric World Simulator for Embodied Interaction Generation Jinkun Hao, Mingda Jia, Ruiyan Wang et al. Source: arXiv cs.CV · Hugging Face: huggingface.co/papers/2604.01001 Date: 1 April 2026 URL: https://arxiv.org/abs/2604.01001
A closed-loop egocentric simulator that generates spatially consistent first-person videos of embodied interactions while dynamically updating 3D scene representations, trained from large-scale egocentric video with an EgoCap smartphone capture system. Relevant to somatic AI precisely because it generates the felt first-person perspective of being-in-motion-in-an-environment—the phenomenological stance central to somatic practice—and models the reciprocal relationship between body and space.
Hao, J. et al. (1 April 2026). EgoSim. arXiv:2604.01001. https://arxiv.org/abs/2604.01001
[11] OmniEgoCap: Camera-Agnostic Egocentric Full-Body Motion Reconstruction 🚩 (original Dec 2025; revised version announced 31 March 2026) Kyungwon Cho, Hanbyul Joo Source: arXiv cs.CV · Date: v2 announced 31 March 2026 URL: https://arxiv.org/abs/2512.19283
A diffusion-based framework reconstructing full-body 3D motion from egocentric wearables (head + hand devices), handling self-occlusion and out-of-sight limbs by processing entire sequences and recovering body-invariant physical attributes (height, proportions). Directly applicable to wearable somatic tracking, first-person movement analysis, and dancer self-monitoring without external cameras.
Cho, K. & Joo, H. (31 March 2026 — v2). OmniEgoCap. arXiv:2512.19283. https://arxiv.org/abs/2512.19283
[12] DreamControl-v2: Autonomous Humanoid Skills via Trainable Guided Diffusion Priors Sudarshan Harithas, Sangkyung Kwak, Pushkal Katara et al. Source: arXiv cs.RO · Date: 31 March 2026 URL: https://arxiv.org/abs/2604.00202
Trains a diffusion model in a unified embodiment space bridging human and robot motion datasets, eliminating manual filtering steps and enabling scalable skill acquisition for the Unitree G1 humanoid. For somatic AI, the cross-embodiment approach advances the idea that movement knowledge can transfer across bodies—with implications for AI dance partners, assistive prosthetics, and somatic simulation.
Harithas, S. et al. (31 March 2026). DreamControl-v2. arXiv:2604.00202. https://arxiv.org/abs/2604.00202
[13] Toward Personalized Movement Coaching: A Data-Driven Framework via Biomechanical Analysis Zhantao Chen, Dongyi He, Jin Fang et al. Source: arXiv cs.CV · Date: 1 April 2026 URL: https://arxiv.org/abs/2604.01130
Markerless motion capture analyses 18 kinematic features (coordination, release mechanics, joint positioning, stability) across 2,396 samples at multiple skill levels, generating personalised reference trajectories using minimum-jerk principles and providing individualized diagnostic feedback. A clear implementation of applied somatic AI: body data → ideal movement model → individual deviation detection → personalised guidance, directly paralleling what a somatic coach or movement therapist does.
Chen, Z. et al. (1 April 2026). Toward Personalized Darts Training. arXiv:2604.01130. https://arxiv.org/abs/2604.01130
[14] BAT: Balancing Agility and Stability for Long-Horizon Whole-Body Humanoid Control Donghoon Baek, Sang-Hun Kim, Sehoon Ha Source: arXiv cs.RO · Date: 1 April 2026 URL: https://arxiv.org/abs/2604.01064
Dynamically switches between two complementary RL controllers (one optimising agility, one stability) using a hierarchical option-aware VQ-VAE, tested on the Unitree G1 humanoid across loco-manipulation tasks. The agility/stability tension mirrors a core somatic concept—the dynamic relationship between release and support, fluidity and structure—making this a conceptually interesting reference point for embodied intelligence in movement.
Baek, D., Kim, S-H., & Ha, S. (1 April 2026). BAT. arXiv:2604.01064. https://arxiv.org/abs/2604.01064
[15] ONE-SHOT: Compositional Human-Environment Video Synthesis Fengyuan Yang, Luying Huang, Jiazhi Guan et al. Source: arXiv cs.CV · Date: 1 April 2026 URL: https://arxiv.org/abs/2604.01043
A parameter-efficient framework for human-centric video generation that separates human motion from environmental elements via cross-attention and a novel Dynamic-Grounded-RoPE positional embedding, enabling extended consistent video without rigid 3D preprocessing. Relevant as a body-to-visual translation pipeline for choreographic visualisation and dance film production from movement inputs.
Yang, F. et al. (1 April 2026). ONE-SHOT. arXiv:2604.01043. https://arxiv.org/abs/2604.01043
[16] Trampoline Gymnastics Pose Estimation via Synthetic MoCap-Derived Data Lea Drolet-Roy, Victor Nogues, Sylvain Gaudet et al. Source: arXiv cs.CV · Date: 1 April 2026 URL: https://arxiv.org/abs/2604.01322
Generates synthetic multi-view training imagery from motion capture to fine-tune ViTPose for extreme, high-speed athletic movement, achieving 19.6% reduction in 3D pose error on real trampoline footage. Demonstrates how MoCap-derived synthetic data can close the gap for unusual or highly expressive movement vocabularies—a methodology directly applicable to somatic and dance datasets where authentic labelled data is scarce.
Drolet-Roy, L. et al. (1 April 2026). Trampoline Gymnastics Pose Estimation. arXiv:2604.01322. https://arxiv.org/abs/2604.01322
[17] LOME: Learning Human-Object Manipulation with Action-Conditioned Egocentric World Model 🚩 (submitted 28 March; featured HuggingFace 3–4 April) Quankai Gao, Jiawei Yang, Qiangeng Xu, Le Chen, Yue Wang Source: arXiv cs.CV / Hugging Face Daily Papers (featured 3–4 April 2026) URL: https://arxiv.org/abs/2603.27449
Generates realistic egocentric videos of human-object interactions conditioned on per-frame body poses and hand gestures, producing physically plausible consequences (e.g., liquid pouring). Advances gesture-conditioned world simulation: given body pose and hand gesture as input, the model generates what the world looks like from within the moving body—directly relevant to AR/VR somatic applications.
Gao, Q. et al. (28 March / featured 3–4 April 2026). LOME. arXiv:2603.27449. https://arxiv.org/abs/2603.27449
Near-window — high field significance:
TokenDance: Token-to-Token Music-to-Dance Generation with Bidirectional Mamba 🚩 (28 March 2026) Ziyue Yang, Kaixing Yang, Xulong Tang URL: https://arxiv.org/abs/2603.27314 — Two-stage music-to-dance generation using Finite Scalar Quantization to discretise both dance and music into tokens, then generating choreography non-autoregressively. State-of-the-art quality and inference speed. Modular token architecture could accept somatic signals as additional conditioning.
PoseDreamer: Scalable Photorealistic Human Data Generation 🚩 (30 March 2026) Lorenza Prospero, Orest Kupyn et al. — Oxford VGG URL: https://arxiv.org/abs/2603.28763 — Generates 500,000+ synthetic labelled 3D human mesh samples via controllable diffusion, achieving parity with real data at 76% improvement in image quality. Directly applicable to bootstrapping somatic body datasets.