Research released from 6–12 October 2026 changed the question from whether a motion system can produce a plausible body to whether it preserves the right correspondence across bodies, sensors, and representations. Three new arXiv preprints show that topology, viewpoint, and latent alignment each require explicit design; a shared architecture alone does not solve them.

CAST corrects motion across unfamiliar skeletons

CAST, submitted 10 October, addresses monocular motion capture when the target skeleton differs from the body in the source video. Learned joint rotations can accumulate error through a kinematic hierarchy, so the system predicts global bone directions and applies reliability-gated corrections from parent to child. Under the paper’s TopoCap protocol, it improves recovery on both familiar and unseen topologies; a causal variant runs above 51 frames per second on one NVIDIA A100.

This is a Tier 3 preprint, not a claim of topology-independent capture in every setting. Its useful contribution is narrower: it identifies which rotational components are observable from bone directions and leaves underconstrained components to the learned predictor.

EgoWholeView makes partial viewpoints cooperate

EgoWholeView, submitted 12 October, uses six synchronised wide-angle RGB cameras on a wearable rig to capture body and two-hand motion without a third-person view. Outward cameras localise the rig; a downward stereo pair estimates body and hand parameters. Reliability signals from the predictions determine how the estimates are fused.

Against monocular prediction, the authors report reductions of 3.7% in body and 11.2% in hand position error after alignment, while temporal refinement lowers angular-acceleration magnitude by about one third. The result matters because portability comes from coordination among imperfect views, not from pretending one view sees everything.

FlowAvatar joins tracking completeness with appearance

FlowAvatar, also submitted 12 October, reconstructs full-body pose, hands, facial expression, eye gaze, and a personalised appearance from sparse inputs on a Meta Quest Pro. The authors report real-time operation and strong temporal consistency on AMASS. Two small user studies—12 and 10 participants—suggest that tracking completeness broadly improves perceived motion quality, while positive overall embodiment requires completeness and personalised appearance together.

The studies do not establish a universal recipe for embodiment. They do demonstrate that visual likeness cannot compensate for missing bodily channels, and complete tracking does not automatically produce self-recognition.

The week’s result: correspondence must be tested directly

Across these papers, “shared” is not a sufficient evaluation. A skeleton mapping must preserve orientation, camera views must agree in a world frame, and an avatar must join tracked action with recognisable appearance. That extends Somatic-AI Lab’s October report on action-relative motion representations and its earlier guide to keeping evidence provenance visible. The practical review question is now: which correspondence was measured, and which was merely assumed?

References

  • Pang, P., Qing, Z., Chen, L.-H., Li, S., Huang, H., & Hu, B. (2026). CAST: Real-time motion capture for any skeleton topology. arXiv. https://arxiv.org/abs/2610.14856
  • Hu, Y., He, X., Cui, D., Shi, Y., & Chen, K. (2026). EgoWholeView: Camera-only head-mounted capture of body and bimanual hand motion for embodied demonstrations. arXiv. https://arxiv.org/abs/2610.16256
  • Gao, C., et al. (2026). FlowAvatar: Real-time full-body avatars from sparse egocentric inputs on consumer XR devices. arXiv. https://arxiv.org/abs/2610.16277