Four releases from 12–18 August 2026 push motion AI beyond the clean 3D skeleton. EEG2MOTION conditions full-body generation on non-invasive brain signals; ReForce transfers human demonstrations to robot hands while tracking force; a plug-and-play interface lets 3D-trained motion-language models accept 2D poses from ordinary video; and HiPHI supplies more than 600 hours of high-precision human motion and object interaction. The shared signal is practical: movement systems are beginning to connect trajectories with intention cues, contact forces, accessible video and physically grounded objects.

Scope note: all four items have date-verifiable public arXiv records in the stated window. The 2D Motion Interface is listed as an ECCV 2026 workshop oral; the other three records are preprints. Dataset sizes and performance statements below are author-reported.

EEG2MOTION conditions full-body generation on non-invasive brain signals

EEG2MOTION: Towards Open-Vocabulary Human Motion Synthesis from Non-invasive Brain Signals — Peng, Pan, Yang, Zheng, Chen, Hu and Zhang; arXiv:2608.14754; 14 August 2026.

EEG2MOTION introduces a dataset of nearly 20,000 paired EEG, motion and text samples spanning thousands of motions. The researchers first align EEG representations with text, video and movement representations, then use an EEG-conditioned masked-motion model to generate continuous full-body sequences.

The work is best read as conditional generation, not literal reconstruction of a hidden movement. Scalp EEG is sparse and noisy relative to detailed body kinematics, so the model learns statistical correspondences from paired examples and supplies much of the output through its motion prior. Even with that caveat, the project widens brain–computer-interface research from small command sets toward a richer movement vocabulary.

For a plain-language account of the distinction between brain signal, inference and generated motion, see the Lab’s explainer on why this is not mind-reading.

→ https://arxiv.org/abs/2608.14754

ReForce transfers contact, not only hand shape

ReForce: Learning Force-aware Retargeting for Dexterous Manipulation — Wu, Zeng, Jing, Ye and Wang; arXiv:2608.15560; 16 August 2026.

Retargeting maps a human demonstration onto a robot with different joints and proportions. Most methods optimise the hand’s configuration, but a visually similar grasp can apply the wrong pressure and fail to move the object. ReForce predicts a residual correction to a kinematically retargeted action so the robot also approaches a desired force pattern. Its general force tracker is trained from large-scale simulated interactions and supports both live teleoperation and offline conversion of demonstrations.

The authors report lower force-tracking error and stronger multi-finger engagement in simulation and on hardware for tasks including paper-cup grasping and using tongs. This is narrow evidence, not a general solution to dexterity. Still, it turns contact force into an explicit transfer target rather than a hoped-for side effect of matching pose.

The distinction extends the Lab’s earlier discussion of the robot that can see a hand but cannot feel it.

→ https://arxiv.org/abs/2608.15560

A 2D interface brings motion-language models closer to ordinary video

A Plug-and-Play 2D Motion Interface for Real-World Motion Language Models — Yokoyama and Ukita; arXiv:2608.15984; 17 August 2026; accepted for an oral presentation at HCMIW, ECCV 2026.

Motion-language models commonly tokenise 3D joint sequences. Recovering reliable 3D motion from a single real-world camera remains difficult, so the authors built an adapter that lets a model pretrained on 3D motion accept 2D input without changing or fine-tuning the original model.

On the evaluated public datasets, the interface performed comparably to 3D input across multiple motion-language models and outperformed models trained from scratch on 2D motion. The team also created a monocular real-video evaluation set. The result does not prove that 2D is intrinsically better than 3D; it shows that under imperfect monocular 3D estimation, a simpler observed representation can be the more useful interface.

→ https://arxiv.org/abs/2608.15984

HiPHI scales high-precision motion and object interaction

HiPHI: A Large-Scale Benchmark for High-Precision Human Motion and Object-Interaction — Ji, Ma, Zhang, et al.; arXiv:2608.16222; 17 August 2026.

HiPHI addresses a familiar trade-off. Internet video offers behavioural range but weak physical state; laboratory capture offers precision but narrow coverage. The new dataset contains more than 600 hours of optical full-body capture with sub-millimetre marker tracking and mesh-level object trajectories. Its collection plan uses FrameNet, a linguistic framework for organising human situations and actions, to broaden coverage of motion and interaction primitives.

The accompanying benchmark examines motion-space diversity, interaction grounding, object consistency and physical-AI applications. Scale and precision are valuable, but neither guarantees representative bodies, cultural practices or lived movement qualities. Those questions depend on sampling and documentation beyond the headline hour count.

→ https://arxiv.org/abs/2608.16222

What to watch next

These releases connect motion to four kinds of context: neural activity, applied force, camera accessibility and object-level physical state. Their usefulness will depend on keeping the provenance of each signal visible. EEG-conditioned motion is generated from learned associations; force is predicted or tracked under a particular embodiment; 2D pose loses depth; optical capture is precise within a designed collection setting. A stronger system can combine these sources, but it should not flatten them into one supposedly complete record of the body.

References

Ji, J., Ma, J., Zhang, R., Yu, R., Wang, W., Chi, W., Peng, Q., Yan, W., Gu, Y., Tian, Y., Wu, T., Li, L., Yuan, C., Dai, R., & Han, L. (2026). HiPHI: A large-scale benchmark for high-precision human motion and object-interaction. arXiv:2608.16222. https://arxiv.org/abs/2608.16222

Peng, Y., Pan, Y., Yang, Y., Zheng, N., Chen, W., Hu, X., & Zhang, S. (2026). EEG2MOTION: Towards open-vocabulary human motion synthesis from non-invasive brain signals. arXiv:2608.14754. https://arxiv.org/abs/2608.14754

Wu, Y., Zeng, L., Jing, C., Ye, J., & Wang, X. (2026). ReForce: Learning force-aware retargeting for dexterous manipulation. arXiv:2608.15560. https://arxiv.org/abs/2608.15560

Yokoyama, K., & Ukita, N. (2026). A plug-and-play 2D motion interface for real-world motion language models. arXiv:2608.15984. https://arxiv.org/abs/2608.15984