Weekly Recap

The knowledge structure, thematic evolution, and publication trajectory of cancer SDT research entered a substantially higher-output phase after 2020, and China had the highest publication count and total citation count and Israel had the highest citations per publication among the ten most productive countries.

This commentary argues that the next implementation challenge for precision oncology is to make clinically meaningful genomic findings realistically actionable, and comprehensive genomic profiling should occur early enough to inform treatment planning.

Through elucidating how electrotherapy reshapes tumor biology and attenuates resistance to established treatments, this review provides new insights into the development of multimodal therapeutic strategies for cancer.

Residence and occupational disparities were associated with FCR and depressive symptoms, and digital health interventions may help address unmet psychological supportive care needs in non-central cities, while vocational rehabilitation support may benefit individuals with professional backgrounds.

The knowledge structure, thematic evolution, and publication trajectory of cancer SDT research entered a substantially higher-output phase after 2020, and China had the highest publication count and total citation count and Israel had the highest citations per publication among the ten most productive countries.

This commentary argues that the next implementation challenge for precision oncology is to make clinically meaningful genomic findings realistically actionable, and comprehensive genomic profiling should occur early enough to inform treatment planning.

Through elucidating how electrotherapy reshapes tumor biology and attenuates resistance to established treatments, this review provides new insights into the development of multimodal therapeutic strategies for cancer.

Residence and occupational disparities were associated with FCR and depressive symptoms, and digital health interventions may help address unmet psychological supportive care needs in non-central cities, while vocational rehabilitation support may benefit individuals with professional backgrounds.

Robots can learn complex dexterous manipulation from just minutes of human "play," outperforming RL by 490x in data efficiency.

Unsupervised environment latents adapt robotic world models to unseen physical properties and rules from just one interaction, drastically improving prediction and task success.

Real-world robot policies can learn autonomously from sparse rewards by treating value predictions as task-progress memory, outperforming offline training by $32.8\%$ success rate.

A single model now masters 3D hand reconstruction, predicts future video under actions, and plans complex manipulation sequences from language, outperforming specialized methods.

General motion tracking policies can be made compliant without sacrificing generality by conditioning a residual adapter on proprioceptive history.

Bench2Drive's dominant comfort metric is fundamentally flawed, rewarding stationary behavior and obscuring the true trade-offs in autonomous driving models.

A specialized CAD harness drastically improves LLM agent performance, cutting output tokens by up to 88% and boosting F1 scores by 13% through geometry-grounded RL.

Robotic harvesting systems can achieve near-perfect sim-to-real transfer for precise, safe manipulation by encoding path-safety and corridor constraints directly into a PPO objective.

Generative robotic policies can now actively select specific behaviors by controlling where generation starts in the source space.

Scene appearance estimation can be solved as a global linear system, slashing computation by orders of magnitude and enabling efficient, high-quality novel-view synthesis.

A simple test-time trigger and latent world model can boost VLA policy success by 15% without retraining, by intelligently selecting the best action from stochastic outputs.

Omnidirectional tiltrotors can now achieve near-perfect wrench feasibility even with slow servo dynamics, outperforming prior methods by over 60% in disturbed maneuvers.

A frozen world-model planner's performance can be drastically improved by adapting its latent rollouts to controller-induced execution shifts, rather than directly substituting predicted trajectories.

Action-free videos unlock significant gains in robot control generalization and robustness by enabling explicit future visual state prediction.

1,000 hours of human touch data unlocks a 37% jump in vision-only dexterous manipulation success, revealing touch as a powerful pretraining signal beyond its direct use in policies.

Lunar drones can now autonomously navigate complex terrain, safely landing even when initial waypoints become unreachable, by leveraging vision-guided optimal control and real-time obstacle avoidance.

Preemptive communication, driven by reachability analysis and risk metrics, slashes inter-agent messages by 77.5% in pursuit-evasion while guaranteeing collision avoidance.

Task-specific visual token compression, not just saliency, unlocks substantial efficiency gains for VLA models without retraining.

Robots can now proactively avoid losing sight of targets by predicting and maneuvering around occlusions, rather than just reacting to them.

Spacecraft inspection planners can be blinded by orbital illumination; a learned VAE measure of image information, not just line-of-sight, dramatically improves inspection quality.

Legal guidelines, not scalar rewards, can guarantee compliant multi-agent motion planning, even with conflicting objectives and private agent goals.

Humanoid robots can now teleoperate effectively with heavy, dynamically changing payloads by online estimating and adapting their full inertial properties.

Lightweight LLMs can infer category-specific risk for real-time UAV navigation, enabling a 3x reduction in collisions and boosting landing success from 34% to 96% in urban environments.

Robots can now react to safety hazards and kinematic limits in under 0.1 milliseconds without interrupting ongoing tasks, shattering previous ROS 2 control limitations.

A fixed-size latent memory, explicitly optimized for robotic control via question-answering and DINO feature reconstruction, unlocks state-of-the-art long-horizon manipulation performance.

Autonomous driving planners achieve 35.8% fewer collisions and 60.8% less lane deviation by directly integrating relational rewards into diffusion model training.

A physics-grounded digital twin for robotic welding achieves sub-millisecond inference and maintains weld quality within 1-3mm root-side-width in simulated reduced-gravity scenarios.

Proactive anomaly anticipation in contact-rich manipulation is achieved by predicting future physical interactions, enabling earlier detection and improved task success.

A complete neuromorphic visual navigation system achieves ANN-level performance in complex environments while slashing power consumption by over 80% for perception and over 87% for navigation.

Fractional-order viscoelastic models are perceptually superior for realistic haptic rendering, outperforming conventional integer-order models in human-subject evaluations.

Robots can now learn complex manipulation from single human videos by extracting object trajectories, bypassing the need for teleoperation and achieving near-perfect task success.

A single policy now masters both pushing and grasping, achieving 57.91% success on unseen objects and enabling complex emergent behaviors like table-assisted reorientation during grasping.

Robots can now distinguish between a user who can't move and one who isn't paying attention, leading to more effective rehabilitation.

Robots can now find and return objects up to 19% faster by treating discovery as a Poisson process and optimizing the entire mission trajectory, including the return trip.

Expert demonstrations reveal a low-dimensional, non-linear geometry in hidden states that, when exploited, significantly boosts imitation learning policy robustness to observation noise.

Robots can now navigate complex indoor spaces without prior mapping by relying on human-seeded topological memories and dynamic, evidence-gated co-updating.

Robot manipulation policies can achieve state-of-the-art performance and 17x faster inference by freezing the VLM and using a lightweight actor to query a fixed task memory.

Robots can now learn complex manipulation skills from human videos by focusing on transferable causal dependencies, bypassing the need for direct motion imitation.

VLMs become practical preference oracles for robot learning, achieving state-of-the-art performance with drastically reduced inference costs and abstention rates.

Industrial safety workers can now be tracked continuously across indoor and outdoor environments with sub-3.5m accuracy using a single, low-power wearable device.

Closed-chain inverse dynamics are uniquely solvable by decomposing motion reconstruction and spatial force balancing, even with indeterminate internal reactions.

A training-free method achieves precise target localization in 3D Gaussian language maps by selectively injecting text priors, doubling mIoU on Replica without sacrificing general queryability.

Global line correspondences unlock robust keypoint matching in challenging, repetitive scenes where local descriptors fail.

Collision-free navigation is achieved with monocular vision alone by directly certifying safety from scale-free relative motion, bypassing explicit obstacle state estimation.

Current VLA models fail to benefit from "How-level" instructions, with enriched training data even decreasing task success.

An attribution gate precisely disentangles user intention from capability degradation in robotic assistance, achieving 3x lower trajectory error than prior methods.

Consolidating diverse VLA skills into one policy is now achievable with a two-level on-policy distillation that leverages student-generated action flows and teacher velocity supervision, outperforming prior methods by up to 18.3% on complex tasks.

VLA policies can now reliably select action prefixes by predicting task completion compatibility, not just generation stability.

A single GPU can now plan complex, contact-rich robotic manipulation trajectories for thousands of parallel worlds, achieving near-80% success on challenging in-hand rotation tasks with unprecedented speed.

Autonomous endovascular navigation achieves 96.9% success on a physical robotic platform using only 2D X-ray fluoroscopy feedback.

Robots can now navigate and manipulate objects based on sound in entirely new environments by fusing acoustic cues with explicit spatial memory derived from egocentric depth observations.

Microbots now navigate complex, dynamic environments with sub-6 µm precision, autonomously avoiding obstacles and delivering cargo.

A novel control strategy effectively prevents motorcycle falls during cornering on slippery surfaces, a long-standing challenge for riders and a major cause of accidents.

Robotic agents can now retain crucial visual-motor history within a compact latent transition space, achieving near-perfect recall on complex manipulation tasks with minimal computational overhead.

Mobile robots can now dynamically place and scale augmented reality projections in real-world environments based solely on language instructions, achieving over 57% success in unseen tasks.

Adaptive scheduling for slow VLA modules can be optimized by a single, measurable quantity, the Value of Intervention, revealing that language-rewrite channels often fail to translate invoked queries into success.

Explicitly aligning video captions with atomic audiovisual events unlocks substantial gains in fine-grained temporal and attributional accuracy, outperforming unstructured generation by over 9% on event-centric metrics.

Execution delays in robotic motion planning can be mitigated by treating them as a distribution of initial states, drastically improving safety and efficiency.

Real-time neural SLAM on a mobile robot is now achievable with state-of-the-art accuracy, bridging the gap between geometric efficiency and neural implicit mapping.

Ablation shock propels micro-LEDs with sub-micrometer precision and 50 $\unicode{x03BC}$m gap tolerance, shattering previous mass transfer bottlenecks.

Verifying complex cyber-physical systems can now be done in days, not years, by combining an ensemble of stopping algorithms with massively parallel HPC simulations, slashing sample requirements by up to 78%.

Learned robotic interaction strategies can be robustly adapted to novel contact conditions by organizing RL policies into generative sequence proposals and refining them with realization-aware manifold updates.

Emotional contagion drives consensus, polarization, and clustering in social networks, with agents converging to agreement based on emotional similarity.

Provable system identification for partially observed linear dynamics is now possible even with up to 50% adversarial data poisoning.

A shared state estimator can enable effective policy-gradient transfer even with differing simulator and plant dynamics, by bounding the gradient discrepancy linearly with model error.

Even state-of-the-art dexterous manipulation policies fail to generalize, achieving only 38.5% success on novel tasks after training on related ones.

Converter-dominated grids exhibit complex, interacting stability phenomena that defy traditional analysis, pushing the boundaries of both model-based and data-driven approaches.

Port-Hamiltonian systems reveal intrinsic limits on energy conversion, with asymptotic efficiencies matching cyclic ones and a generalized Carnot bound emerging.

CBF safety filters can create unstable periodic orbits and homoclinic loops, even for stable linear systems, fundamentally altering system dynamics and stability.

Egocentric human videos can now directly train humanoid robots, bypassing the need for expensive robot-specific demonstrations for novel tasks.

Drillstring-wellbore friction, often overlooked, critically impacts the stability margins of offshore drilling systems, even when well-posedness is maintained.

eVTOL vertiports can unlock significantly more departure directions by leveraging aircraft climb capabilities against urban obstacles and prevailing winds.

A novel flow matching approach bypasses iterative rendering and complex 2D-3D registration for robust monocular robot pose estimation, achieving superior zero-shot cross-scene generalization.

Independent antenna repositioning can boost underground mine communication link reliability by up to 22 dB, overcoming severe spatial fading.

Diffusion models can achieve up to $6.9\times$ speedup by adaptively adjusting step sizes based on a staleness signal's "headroom," effectively jumping over scheduled inference steps.

A single feed-forward network reconstructs dense 4D human correspondence, including clothing and hair, from monocular RGB images without a body template, achieving amodal tracking and precise surface point localization.

Maximizing the determinant of an RBF Gram matrix on initial noise vectors unlocks collective diversity in diffusion models, outperforming pairwise repulsion and boosting downstream few-shot classification accuracy.

Event-triggered data selection slashes communication costs for online system identification without sacrificing accuracy, enabling real-time monitoring of dynamic systems.

Battery degradation can be accurately predicted and understood from raw operational data alone, bypassing the need for engineered features or limited historical windows.

Learned relationships from healthy sensor data enable highly accurate diagnosis and significant fuel correction, reducing CO2 emissions accounting error by over 80% in heavy-duty diesel vehicles.

AI agents can now generate and verify complex 3D worlds using executable code, outperforming leading models in human preference studies.

Achieving reliable physical autonomy demands a system-level perspective, integrating diverse components beyond isolated model performance to navigate real-world complexities.

Autoregressive video generation can be sped up by over 1.5x without sacrificing quality by intelligently freezing tokens that have already converged.

Multi-turn image editing models can now maintain high fidelity and instruction adherence over dozens of edits by training them to robustly edit their own prior outputs.

Open-set, person-specific video descriptions are now achievable with grounded spatial information, outperforming SOTA VLMs and introducing a more robust evaluation framework.

Semantic structure, not just token-level feedback, unlocks significant gains in RL-driven text-to-image prompt engineering.

Diffusion model token merging can be optimized by analyzing denoising dynamics, not just prompt relevance, to preserve structural coherence and visual fidelity.

Even state-of-the-art MLLMs fail to match human accuracy in verifying 3D content, highlighting a critical gap in grounding generative AI.

Retinal vessel segmentation now achieves near-supervised performance without any manual labels, unlocking large-scale analysis of retinal images.

Remote KV caches can cripple LLM prefill latency, but sharding the KV store across parallel ranks and enabling asynchronous loading slashes p99 time from 83s to under 1s.

Geometry becomes position: a novel Dual-Kernel Attention mechanism injects 3D scene geometry directly into vision-language model attention, bypassing challenging feature fusion and achieving SOTA spatial reasoning.

Existing audio-video generation metrics fail complex scenes; this new event-centric framework achieves state-of-the-art human alignment and pinpoints generation failures.

Adaptive uncertainty measurement, not just prediction quality, is the critical factor for robust multi-object tracking, enabling a +3.1 HOTA improvement by dynamically adjusting search regions.

Unsupervised clustering of public transport vehicles reveals distinct routes and types by fusing deep morphological features with color histograms, outperforming morphology-only approaches.

Wavelet decomposition unlocks LLMs for lossless image compression, achieving state-of-the-art rates by transforming heterogeneous signals into structured subproblems.