The 26 August–1 September scan found two directly relevant research releases and one evaluation announcement worth watching. PoseOFF anchors optical flow around joints to anticipate action earlier. IMPACT uses an internal interaction map to focus world-model training on changing regions. Google DeepMind’s double-blind evaluation pilot is not a motion paper, but its contamination-control idea is relevant to any benchmark claiming progress in embodied behaviour.

PoseOFF anticipates before an action is complete

Grundy, McCarthy and Fluke’s Pose-Anchored Optical Flow for Low-Latency Human Action Anticipation in Human-Robot Teaming (arXiv:2608.25495, 26 August 2026) conditions local flow on pose. The authors report gains at earlier observation ratios, suggesting a robot can act on less of a sequence without processing every pixel equally.

→ https://arxiv.org/abs/2608.25495

IMPACT trains on the interaction map

Tang and colleagues’ IMPACT (arXiv:2609.00161, 31 August 2026) identifies a supervision mismatch in global denoising loss: static content dominates while sparse hand–object changes are underweighted. Attention around object tokens supplies a prior for reweighting local prediction errors. The reported gains cover robot-arm and human-hand manipulation, with physical plausibility evaluated alongside visual quality.

→ https://arxiv.org/abs/2609.00161

Double-blind evaluation targets benchmark contamination

Google DeepMind announced a pilot for cryptographically isolated external evaluation on 27 August 2026. The stated goal is to prevent a proprietary model from seeing test questions before assessment. The release concerns general AI evaluation, not motion, but the principle transfers: if a motion model or robot has indirectly encountered a benchmark sequence, a high score says less about generalisation.

→ https://deepmind.google/blog/piloting-the-worlds-first-double-blind-ai-evaluations/

What to watch next

These releases place scrutiny at three points: before an action is complete, where an interaction changes, and before a benchmark result is trusted. For movement research, those are connected. Early anticipation can be useful only if the signal is not leaking future frames; interaction maps matter only if the changing region is correctly identified; and benchmark gains matter only if the test remains genuinely held out. The archive’s contact-sensitive tracking report and motion-evaluation guide offer the relevant safeguards.

References

Google DeepMind. (2026, August 27). Piloting the world’s first double-blind AI evaluations. https://deepmind.google/blog/piloting-the-worlds-first-double-blind-ai-evaluations/

Grundy, L. de Z., McCarthy, C., & Fluke, C. (2026). Pose-anchored optical flow for low-latency human action anticipation in human-robot teaming. arXiv:2608.25495. https://arxiv.org/abs/2608.25495

Tang, R., Fang, J., Wang, Z., et al. (2026). IMPACT: Attention is the interaction map for scalable interaction-aware world model training. arXiv:2609.00161. https://arxiv.org/abs/2609.00161