The Handle Stays Close

A new preprint uses operando charge photometry, cryogenic microscopy and modeling to show connected water-rich pathways inside hydrated organic mixed ionic-electronic conductors. Changing one polymer’s terminal side chain then produced two different device behaviors: the hydrophobic OTBS version showed about 0.3% passive swelling, retained roughly 0.53 V after 30 minutes versus about 0.38 V for a hydrophilic control, and supported 100 gradual conductance states; the hydrophilic material switched in about 1 ms and amplified an ECG waveform. These are laboratory demonstrations, not evidence of long-term cycling, manufacturability, biocompatibility or in-vivo performance.

Dex-One2Many converts one human demonstration video into stage-wise scene graphs that describe object relations and order rather than copying a single trajectory. Those relations generate varied reset states and predicate-based rewards for simulation training before zero-shot transfer to a UR3 arm with a 20-degree-of-freedom hand. The protocol assigns 30 real-world trials per task and method—10 seen and 20 unseen—although one baseline task was unavailable after a hardware failure. Across five tasks, the paper reports 60–85% success on unseen configurations and a 75% mean, while its baselines reached no more than 15%. The evidence remains bounded to one robot setup and depends on vision-language extraction, pose estimation and predefined predicates.

Hybrid Cinematography pairs an editable 3D shot plan with a lightweight proxy of the recorded take, then previews where a requested camera move would depend on content the camera never captured. A mobile AR workflow lets filmmakers change the path or timing, record targeted pickups, or knowingly accept synthesis before a slow generative pass begins. In a qualitative study, seven experienced filmmakers rated the preview a median 7 of 7 for showing where and why a model must invent content, and a median 6 of 7 for exploring trade-offs. The preview measures physical support rather than final model uncertainty; the study has no comparison condition and current mobile processing can thermally throttle.

SpaceFlow is a training-free 3D-generation pipeline that assigns a separate control strength and optional appearance cue to each geometric primitive in an object. One part can hold close to supplied geometry while another remains open to generative variation, replacing a single global compromise with local decisions. Across 83 assets, the authors report stronger regional separation without reducing their CLIP measure; in 337 pairwise trials from 29 evaluators, SpaceFlow was preferred to a strict high-guidance baseline 71% of the time and remained competitive with low guidance. Conflicting text and geometry can still distort shapes, and coarse routing on a 32×32×32 voxel grid can bleed appearance across material boundaries.

A peer-reviewed preclinical study places a deterministic state machine beneath a multimodal TIAGo rehabilitation workflow. The robot handles predefined low-risk phases, while malaise, stop requests, repeated refusal or unresolved speech events move the system into a waiting state for a therapist to resume, repeat, skip or terminate. Sixteen healthy participants completed two sessions one month apart without a safety-related interruption; mean SUS scores were 73.75 and 79.69. Suspicious dialogue transcripts still appeared in 14.2% and 17.1% of 240 interactions. This is technical feasibility, not clinical efficacy: no stroke patients participated, failures were not systematically injected and emergency-stop latency was not measured.