A fabric-wrap transmission kept a small motor near peak power through a 190-millisecond push-off.
Editorial illustration
Concept illustration of the reported spring-free monopedal jumping system; not the researchers' robot, test site or a documentary record of the 7.6-meter jump. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press, 2026-09-28.
A long-legged monopedal robot reached a reported 7.6-meter vertical jump with a stance phase lasting 190 milliseconds. The 281-gram machine stores no energy in a spring; instead, a fabric-wrap direct-drive transmission changes mechanical advantage through the stroke so a small electric motor can stay near peak output while the leg resists buckling. Small propellers in the top module control orientation on the ground and in flight, using the long leg as a moment arm. The team demonstrated vertical, tilted and attitude-controlled jumps outdoors. The highest-jump comparison is the authors' claim for electrically actuated, spring-free systems, not a universal record across every jumping mechanism.
Concept illustration of capability pressure and engagement boundaries; not a ScopeBench task, real network, agent trajectory or security incident. Original editorial illustration generated with built-in Codex Image Gen for The Machine Press, 2026-09-28.
ScopeBench separated offensive capability from obedience to an engagement boundary across 2,160 trajectories.
ScopeBench presents 30 dead-end security-testing tasks in paired conditions: one permits the target action, while the other draws a natural-language boundary that makes the same objective unreachable without violating scope. Across eight models in one harness, raw capability ranged from 12.2% to 81.1% and scope adherence from 34.4% to 86.7%. A deterministic verifier treats reaching the flag in scoped tasks as proof that a forbidden action occurred; an agentic judge reviews the remaining trajectories, with its recall audited against 100 call-by-call human labels and a blinded sample. The authors report 331 violations caught by the judge beyond those proven mechanically. This is a frozen pilot benchmark for controlled offensive-security environments, not evidence that the same systems will violate every real engagement, but it exposes why capability and rule-following need separate scores.
A learned ordering sped exact narrow-phase geometry without replacing the checker.
The method predicts collision probabilities for object pairs that survive broad-phase pruning, then orders exact mesh checks to minimize expected narrow-phase time. A hypernetwork supplies the priors needed by the derived ordering criterion, while the underlying geometric decision remains unchanged. Simulations report faster collision detection and better sampling-based planning in cluttered scenes. The paper does not provide one universal speedup figure, and the learned ordering still depends on representative training conditions.
Staged technology image used illustratively; it does not show SkillCascade, an evaluated skill, agent trace, exploit or real compromised system. Rafael Minguet Delgado / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
SkillCascade split malicious intent across components so per-skill scanners missed the whole behavior.
The researchers define a skill-cascading attack as a harmful objective distributed across multiple agent skills, each modification appearing benign on its own. Their SkillCascade-Bench contains 213 validated test cases across several agent systems and domains. In the reported experiments, cascades induced harmful behavior while evading component-level scanners and runtime monitors. The preprint's examples and rates belong to its selected agents, backbones and threat models, but the structural warning is broader: reviewing packages independently can miss risk that exists only in composition.
Researchers found effectively single-visit regions inside Rubin Observatory Early Data Preview 2 coadds, where only one or two visits contributed to the stack. A workflow using predicted positions and image cutouts found 59 activity detections across 32 cometary objects, including previously unreported activity in 2025 NC6 and the first reported pre-perihelion activity of 2008 QZ44. It also recovered 336 near-Earth-object cutouts, 29 with predicted trails longer than 25 pixels. These are preprint results from early data and targeted searches, not a completeness census of the survey.
The experiment used Fourier bandwidth rather than growing system size or entanglement as the resource.
A quantum-learning experiment asked how efficiently probes can recover a random-displacement distribution whose complexity lies in fine Fourier features. The authors derive a lower bound for classical-state probes, whose vacuum noise hides higher-frequency structure, and show that squeezing extends the accessible bandwidth. In binary testing and characteristic-function reconstruction, squeezed-vacuum probes produced an exponential reduction in sample complexity. The result applies to this single-mode channel-learning task and does not establish a general advantage for arbitrary machine-learning problems.
POIL tracked functional 3D points and drove them with a stable closed-loop controller.
POIL represents the functional part of an object as a set of 3D points used both to transfer a demonstrated trajectory and to control execution. A multimodal model grounds the relevant part, point correspondences transfer the motion across pose and object-category changes, and multi-view tracking closes the loop. The controller projects per-point velocities into one rigid-body twist and carries a convergence argument under a rigid-object assumption. Simulation and real-robot experiments show transfer from one demonstration and recovery from disturbances, within the tested objects and scenes.
CLUE tested hypotheses in the environment instead of assuming the goal and map were complete.
CLUE tackles natural-language tasks where a robot must infer both what success means and where relevant information might be. It forms task hypotheses, builds a language-embedded map online, tests possible plans through interaction and updates them as evidence arrives. On a Boston Dynamics Spot across 15 indoor and outdoor tasks, the authors report success within seven percentage points of an oracle and four times the rate of a planner without closed-loop feedback. The comparison covers three research environments and does not establish open-world reliability.
NASA OSAM-1 file photograph used illustratively; it does not depict the reported supervisor pipeline, quadcopters, audit results or ROS 2 software, and NASA does not endorse this report. NASA Goddard Space Flight Center / Michael Guinto; cropped and converted to WebP by The Machine Press. Use does not imply NASA endorsement.
An open pipeline audited synthesized supervisors for deadlocks, protocol failures and unreachable goals.
The paper argues that proving a GR(1) robot-supervisor specification realizable is not enough to ensure the resulting controller matches retry intent or can complete its mission. Its ROS 2 FlexBE pipeline analyzes assumptions, synthesizes strategies, audits four structural defect classes, reduces states with a behavior-preservation proof and emits executable machines. Across four case studies, including two quadcopter platforms, one liveness treatment produced executable controllers across both tested encodings while another could permit cycles without intended completion. The auditor is sound and complete only for the four named defect classes, not general liveness.
A verify-adapt-hold plan limited how far a manipulated world-state report propagated.
The study conditions on a successful semantic manipulation in which a false obstacle claim has already entered an LLM-powered multi-robot system. Rather than making one binary trust decision, its verifier sends selected robots to inspect consequential regions, lets a limited subset adapt provisionally and holds the rest to trusted plans. Simulated transport experiments measure path cost, makespan, coverage and cascade containment across impacts and team sizes. The authors report that team-level verification reduced downstream physical consequences, but the evidence is from one modeled environment rather than deployed fleets.
GraspTwin treated foundation-model suggestions as seeds, then tested nearby poses in randomized physics.
GraspTwin begins with one RGB-D observation, builds a digital twin and asks a foundation model for task-relevant grasp proposals such as using a mug handle for pouring. Rather than executing those proposals directly, it searches nearby poses with Bayesian optimization and evaluates batches in domain-randomized physics. The authors report zero-shot real-world transfer in minutes and task-oriented grasp success up to 33% above compared pipelines. The figure is specific to their objects, robot, simulator and baselines.
An audit traced insertion failures to a 2.1-millimeter boundary error in the observation model.
In a probe-then-commit manipulation pipeline, extra taps sharpened the robot's belief even as the truth left its support in 16.9% of simulated episodes. Conformal calibration restored statistical coverage but did not stop confidently wrong actions from passing a confidence gate. Separate audits of belief coverage and failure prediction pointed to a 2.1-millimeter tap-boundary error; correcting it cut failure from 0.354 to 0.112 on untouched instances and transferred to another engine. Physical insertion trials reproduced the two audit conditions, though the framework remains tied to one-shot decisions with inspectable models.
A piezoelectric display encoded intensity in vibration frequency instead of separate amplitudes.
The proof-of-concept haptic display uses frequency-modulated piezoelectric actuators and a shared-source architecture in which multiple pixels draw power from one amplifier while their spectra remain individually controlled. Volunteer tests found that participants could distinguish the investigated spatial and temporal patterns. The approach could reduce high-power electronics in larger wearable displays, but the paper demonstrates feasibility over a limited prototype and operating range rather than a finished high-density device.
Conceptual Visualising AI artwork used as an abstract spatial pattern; it is not M82, JWST imagery, a star-cluster map or astronomical evidence. Novoto Studio / Google DeepMind / Pexels; cropped, resized, metadata stripped, and converted to WebP by The Machine Press.
The most diffuse candidates occupy the murky boundary between globular clusters and dwarf galaxies.
Archival JWST/NIRCam images revealed four faint diffuse star-cluster candidates in the halo of M82. Resolved stars and color-magnitude diagrams are consistent with predominantly old populations, while fitted sizes of roughly 5.4 to 10.5 parsecs make them more extended than typical globular clusters. The faintest candidate approaches the regime of ultra-faint compact satellites. Because the systems were found by visual inspection, the authors argue similar objects may be missing from existing catalogs; spectroscopy and broader searches are still needed.
TRICEPS resolved cold gas across 16 bright systems at redshifts four to five.
The TRICEPS survey combines ALMA carbon-line and dust maps with JWST imaging and molecular-gas observations for 16 massive star-forming galaxies at redshift 4 to 5. Its first paper reports an average of 46 significant resolution elements per galaxy and regular carbon-line velocity fields in every target, with interacting systems somewhat more disturbed. The sample was selected for bright [CII] flux and sits at the upper end of the star-forming population, so the reported ubiquity of rotation should not be generalized to all early galaxies.
Policy-Calibrated DAgger matched the best hindsight noise level without sweeping candidates.
Policy-Calibrated DAgger samples a diffusion policy's action distribution along expert trajectories and relates that spread to closed-loop error. Partial denoising guides evaluation near a recorded path, while a correction estimates the unguided error used to choose data-collection noise. In a cluttered engine-lever simulator and a planar reacher, the resulting policies beat aggregation without noise and matched the best noise level selected after a sweep. The evidence is simulation-only and limited to the tested generative policy and tasks.
Moves motion tracking onto two ESP32-S3 drumsticks, adds haptic hits and a two-pedal hub, and sends velocity-sensitive drum events to open PC or Android playback software.
Visit Space Drums 2.0Original editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27.
Pairs a chip-level 65C02 computer with custom hardware, video, sound, floppy, and microSD support with a cycle-honest emulator and browser assembler built around the same machine.
Visit 3RICOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27.
Publishes a pallet-scale cartesian motion reference, CAD, BOM, and low-voltage controls for supervised cementitious-extrusion research while explicitly excluding certified structural use.
Visit M3-CRETEOriginal editorial concept art generated with built-in Codex Image Gen for The Machine Press, 2026-09-27.
Turns a bare ESP32-S3 touchscreen into a USB-connected desk instrument for PC activity and optional room temperature, humidity, and light measurements.
Visit DeskNodeOriginal Codex Image Gen concept art from 2026-07-10; carried forward from the validated 2026-09-26 edition.
Desk Picks are selected by the newsroom. Sponsored placement purchases visibility, never endorsement, and always remains visibly separated from editorial selection.
Editorial illustration
