EgoAERO is a novel framework that learns dexterous robot manipulation from a single egocentric RGB-D human demonstration without requiring pre-scanned object assets. It reconstructs contact-consistent hand-object trajectories and converts them into robot policies using two-stage residual learning. The framework also introduces an online quality assessment mechanism and compiles the large-scale EgoDex-R dataset. Experiments show it achieves performance close to CAD-based methods.

MuJoCo-Drones-Gym is an open-source Gymnasium-compatible multi-drone environment built on MuJoCo, supporting arbitrary Crazyflie 2.x nano-quadcopters with modular physics, action, and observation APIs, and including a PettingZoo wrapper for multi-agent RL and seven task environments.

IntentNav is a framework that learns human-like object navigation policies from human demonstrations. It uses frontier-based human-intent labeling to infer high-level search intent and constructs a spatial-visual candidate space for efficient exploration. It achieves state-of-the-art performance on multiple benchmarks and zero-shot transfers to various robot platforms.

Q-VGM is an off-policy RL method that efficiently fine-tunes flow-matching VLA policies by matching value gradients through a denoising-time field, avoiding backpropagation issues. It achieves significant success rate improvements on multiple benchmarks and real-robot tasks.

PRISM is a task-agnostic framework that extracts a state-conditioned Gaussian prior from the world model's encoder via a lightweight MLP, and fuses it into the planner's sampling distribution using a precision-weighted Product-of-Gaussians update, significantly improving success rates in continuous control tasks.

This paper presents X-OP, a hierarchical whole-body teleoperation framework using a single XR device that generalizes across robot morphologies without retraining. It features an MPC-based motion retargeter jointly optimizing operator intent and dynamic feasibility, with state synchronization and SLAM feedback for robust execution. Simulations show over 30% faster task completion and 20% lower power consumption for a humanoid, and zero collisions for a mobile manipulator. Real-world experiments confirm effectiveness and user adaptability. The plug-and-play framework enables scalable, morphology-agnostic teleoperation with real-time behavior customization.

Researchers developed a real-time end-to-end controller for powered knee-ankle prostheses using Temporal Convolutional Networks, eliminating need for manual tuning and intent classifiers. Tested on level ground, ramps, and stairs, the controller demonstrated seamless adaptation, matching training-data scaling and enabling natural stair transitions without subject-specific adjustments.

This paper proposes a path planning method for autonomous vehicles in threat-laden environments using Deep Deterministic Policy Gradient (DDPG). Threats are modeled as circular no-go zones. The DDPG agent learns a direct mapping from state to actions via trial and error, with a reward function comprising attractive, repulsive, and energy penalty components. Compared to traditional pseudospectral optimal control, DDPG is significantly faster while producing effective paths, making it suitable for real-time applications.

MinNav is a navigation stack based on optical flow and its uncertainty, enabling tiny aerial robots to fly through scenes with static/dynamic obstacles and unknown-shaped gaps without prior knowledge. It achieves a 70% success rate in real-world experiments and matches depth-based methods with far less computation.

This paper presents VoLoAgent, a VLM-based system for open-vocabulary long-horizon manipulation that orchestrates heterogeneous robot capabilities as interruptible tools. It introduces a closed agent loop for planning, monitoring, and recovery in real physical environments. The RoboVoLo benchmark is also introduced, demonstrating significant improvements over single VLA/VLM or tool-based systems.

NeuroAlign is a hierarchical multimodal fusion framework that integrates fMRI and DTI data for the analysis of Mild Cognitive Impairment (MCI). It introduces Dual-Modal Hierarchical Alignment (DMHA) and Dual-Domain Hierarchical Interaction (DDHI), along with a gradient-free attribution method called Synergistic Activation Mapping (SAM), achieving competitive detection performance on multiple datasets.

AMN (Adaptive Multi-Scale Nuclei Network) is a dual-encoder segmentation framework that combines Swin Transformer and ResNet-50 feature pyramid via a learned per-channel gating mechanism. It achieves mean Dice 0.82 and mean F1 0.68 on CoNIC benchmark, outperforming eight baselines, and shows strong generalization on MoNuSeg.

This paper addresses the challenge of perception in dense, unstructured urban traffic for autonomous driving by proposing a 360-degree LiDAR perception framework that combines sector-wise panoramic processing with rotation-equivariant sparse convolutions. Evaluated on a custom dataset from diverse Indian traffic conditions, the framework achieves strong detection for cars (92.02/90.51), buses (80.53/76.34), and trucks (78.59/74.16), while lower performance on pedestrians (67.45/61.02), cyclists (73.21/69.54), and motorcyclists (71.20/68.13) highlights the difficulty of detecting smaller, variable road users.

This paper proposes a statistical fault injection framework named SENTRY that leverages finite-population sampling theory to provide formal reliability guarantees for Vision Transformers. With only a few thousand samples, it bounds failure rate estimation within 1% error margin at 99% confidence, achieving up to 10,700x reduction in experimental cost. The study reveals a highly non-uniform reliability landscape: only 3% of FP32 bit-flips cause failure, but most of those lead to catastrophic accuracy collapse, with vulnerabilities localized to normalization layers and critical exponent bits in IEEE-754 format.

A new study tests humans and advanced multimodal AI models on distinguishing real legal photos from AI-generated fakes. Results show humans average 64.8% accuracy, dropping to near chance on top generators, while AI models never misidentify real images but miss most synthetic ones. The authors advocate for a combined approach of trained human review, AI screening, and provenance infrastructure like C2PA.

This paper introduces VisualLeakBench, a benchmark for evaluating action-boundary propagation failures in vision-language agents, where sensitive text from images is copied into tool arguments. Baseline tests show 78.8% PII and 85.5% unsafe-text propagation rates. Defensive prompts reduce PII to 2.0% but suppress utility, while unsafe-text remains at 52.6%.

Researchers at MIT analyzed adversarial fine-tuning of Vision Transformers (ViTs) to improve robustness to image perturbations. They found that fine-tuning on specific corruptions (low/high frequency) improves performance on those corruptions but does not generalize to unseen types. Despite changes in attention and knowledge, sparse representations remained unchanged.

SlideCheck is a lightweight tool that uses frozen patch features from pathology foundation models to score abnormality and malignancy in whole-slide images, enabling better control over pretraining data composition. Experiments show that data distributions defined by SlideCheck influence downstream ViT pretraining performance, and curated subsets can match full-data results.

This doctoral thesis proposes two privacy-preserving frameworks for group emotion recognition (GER) using collective audio-video signals instead of individual cues. The first uses cross-attention multimodal fusion with Frames Attention Pooling (FAP), the second (VE-MD) learns a shared latent space for emotion classification and structural representation prediction. Results show competitive performance without individual features.

A page image classifier fine-tuned on over 48,000 annotated historical document images achieves 99.16% accuracy using RegNetY-16GF, enabling automated sorting for OCR and data extraction. CLIP models underperform on unlabeled data despite high test accuracy.

This paper extends MST-Direct, a Sinkhorn optimal transport approach, to multivariate, conditional, and large-scale geostatistical simulation. It addresses scalability via a sparse candidate-restricted Sinkhorn matcher with O(nC) memory, extends to multiple variables using an FFT-MA Gaussian backbone, and enables hard-data conditioning through kriging. Validation on a six-variate heteroscedastic distribution shows zero histogram error and exact spatial correlation reproduction, outperforming the PPMT approximation.

This paper introduces CARTOGRAPH, a verification layer for AI scientists that couples unresolved-subspace experiment steering, explicit ambiguity closure, and residual-based library inadequacy detection. It outperforms raw projection in five testbeds and demonstrates the ability to tentatively identify and then revoke out-of-library mechanisms. In a retrospective audit, it flagged all inconclusive claims from the A-Lab autonomous materials system while passing most confirmed claims.

Diffusion Language Models (DLMs) suffer from bidirectional attention causing failure of existing KV caching methods, leading to near-zero accuracy. The proposed bidirectional prefix caching (bicache) dynamically identifies safe layer depths to reuse shared prefix KVs, improving throughput by 36.3%-98.3% with only 0-1.8% accuracy degradation.

Accurate carbon emission monitoring is critical for climate policy, yet high-frequency city-level data is scarce. TriHead-GAN is a Transformer-based adversarial framework whose triple-head discriminator jointly supervises distributional authenticity, cross-variable dependency, and temporal smoothness. Experiments show it outperforms baselines and improves downstream forecasting in low-resource settings.

STARIXNet is a lightweight neural network that guides resource allocation in multivariate space by capturing spatio-temporal relationships among multiple system metrics, prioritizing service stability and cost efficiency. Deployed at Walmart for critical microservices, it achieves 10-50% cost savings and improves service stability and customer experience.

A new Hierarchical Emergence Framework (HEF) models emergence as phase transitions in mechanism landscapes, explaining convergence across machine learning, biology, and physics. Experiments on modular arithmetic transformers show weight norm peaks precede grokking in 92% of runs, and all grokked models converge to 0.9745±0.014 accuracy, supporting the framework's predictions.

The SPIN framework tackles exponential action space and communication overhead in decentralized swarm control by modeling swarm topologies as compressed tensor networks, reducing complexity from exponential to linear.

Researchers introduce MedicalRec, a transformer-based recommender system that recommends optimal models for medical image classification without retraining, reducing energy consumption and trial-and-error waste. The system uses a benchmark dataset MedicalRec-Bench with over 5,000 records from 3,000 articles, achieving a maximum HitRate@100 of 75.5%.