A Star Trek analogy about AI safety: knowing the right strategy is not enough; execution matters. The quote highlights the gap between intention and action in AI systems.

We present the stochastic decoupled policy gradient (SDPG), a lightweight visual reinforcement learning method that trains diverse visuomotor control policies end-to-end within a few hours on a single NVIDIA RTX 4080 GPU. SDPG estimates policy gradients via random perturbations of trajectory rollouts, requiring orders of magnitude fewer batch-rendered environments and substantially reducing compute and memory overhead. On visual MuJoCo benchmarks, SDPG consistently outperforms baseline methods in training time, memory usage, and rewards. Finally, we introduce a suite of realistic visual robotics benchmarks spanning dexterous manipulation, challenging locomotion, and demonstrate effective sim-to-real transfer on physical hardware.

In dynamic urban logistics, stochastic time-sensitive tasks challenge heterogeneous AAVs task allocation optimality. This paper proposes a RL-enhanced overlapping coalition formation game, using a transformer-based soft actor-critic network to adapt to time-varying task sets. Numerical simulations show a 39.76% cost reduction; indoor flights validate practicality.

This paper proposes Robust Koopman-CBF SAC, a safety-filtered actor-critic framework that learns a Koopman predictor from data, constructs affine CBF constraints in a lifted space, and enforces them via a quadratic-program safety layer with robustness to approximation error. It achieves zero constraint violations on CartPole benchmarks while matching or exceeding unconstrained SAC returns, but reveals limitations on high-dimensional tasks.

This paper presents a decentralized approach called R2P2 for collaborative box transport by multiple robots across flat, uphill, and downhill terrains with varying friction. Robots are assigned roles (push, support, prevent) based on rules and use proportional velocity control, reducing communication and synchronization needs. Evaluated in simulation with six robots and validated physically with four turtlebots moving a 1.2 kg box, R2P2 outperforms virtual-leader-follower methods in success rate.

Teleoperation is key for robot data collection, but novices often produce suboptimal demonstrations. The DQAF framework provides immediate post-episode feedback to improve quality.

RCSP is a predictive planning layer that addresses the near-miss commitment problem in mobile robot navigation by evaluating candidate commands against plausible short-horizon obstacle futures. Simulations show it enhances safety and path quality but adds latency, revealing its role as a complementary module for existing navigation stacks.

NightSight presents a lightweight perception approach combining a monocular event camera, coded aperture lens, and IR dot projector to enable autonomous navigation in complete darkness for small aerial robots. The system uses depth-dependent blur from the coded aperture to train a CNN on synthetic data, achieving zero-shot generalization to real scenes. It runs at 20 Hz on an NVIDIA Jetson Orin Nano with 7.0 cm error up to 2.5 m range.

A framework for heterogeneous robot collaboration under bandwidth constraints, using β-Sparse Gaussian Processes for task-aware point selection and balancing exploration, achieving 18% path cost reduction and 76% information reduction in simulations.

This paper introduces PhyPush, a physics-guided Transformer framework that estimates an object's mass and friction coefficient using only end-effector velocity from a single push, eliminating the need for force/torque sensors. Experiments show reduced error in simulation and real-world settings.

This paper proposes Belief-Aware GSAC (BA-GSAC), which adaptively modulates the distillation coefficient λ via ensemble disagreement, and systematically investigates when adaptive guidance is beneficial for autonomous driving under partial observability. Experiments show benefits under mild to moderate occlusion, but under severe occlusion the adaptive coefficient collapses due to 'observability blindness'—the ensemble predicts partial observations and fails to detect missing information. Proposed fix: train ensemble on full-state predictions. Simple linear decay schedule outperforms adaptive methods, indicating stability gain stems from scheduling effect.

This study benchmarks 12 architectures across four model families on the Retinal Fundus Multi-disease Image Dataset (RFMiD) for binary screening and multi-label classification. All models achieve AUC>84% in binary screening, with attention-based models (SwinTiny, CoAtNet0, MaxViTTiny) performing best. Vision-language models are competitive with CNNs but do not surpass top transformers and hybrids. External validation on Messidor-2 yields AUC 66.8%-84.7%, with hybrid and transformer models demonstrating strong performance.

VesselSim proposes a two-stage framework that generates 16,500 anatomically plausible 3D angiographic volumes via stochastic geometry-driven simulation and trains a 3D U-Net solely on synthetic data. Test-time adaptation via a self-supervised mask reconstruction decoder bridges the domain gap, achieving competitive performance with state-of-the-art on real MRI and CT datasets, significantly reducing reliance on expert annotations.

Researchers propose Dimensional Distribution Emotion State (DDES), a new emotion representation using valence and arousal to predict emotional responses to artworks, aiding museum curators in designing emotion-based exhibitions.

LongAV-Compass is a systematic benchmark for evaluating minute-long audio-visual generation across text, image, and video conditioning. It contains 284 test cases, integrates MLLM-assisted assessment with perceptual metrics, and evaluates over 20 dimensions.

RoMo is a large-scale, high-quality human motion dataset that addresses the trade-off between small mocap datasets and large low-quality in-the-wild collections. It uses a taxonomy-aware filtering pipeline, a three-level semantic taxonomy for annotation, and a fine-grained evaluation framework. Models trained on RoMo achieve state-of-the-art fidelity and diversity, and the accompanying Motion Toolbox standardizes metrics and data conversion.

This paper studies cooperative spatial intelligence for decentralized embodied agents in city-scale outdoor environments, introducing the Sentinel Challenge benchmark and the CoSaR framework that combines foundation model communication with classical navigation algorithms, leading to faster gathering and improved safety.

DuoGesture is a neuro-inspired, biomechanically informed dual-stream approach that decomposes co-speech gesture synthesis into coupled semantic and beat streams. It uses a Semantic Variational Information Bottleneck for coordination, Motion-Grounded Semantic Conditioning for the semantic stream, and an Inertial Beat Prior for the beat stream. Experiments show it outperforms holistic baselines.

Pre-trained video LLMs struggle with auxiliary streams like audio or depth maps due to modality interference. UniMVU uses instruction-aware dynamic gating at two levels (inner-modality and modality-level) to adaptively balance importance, achieving gains up to 13.5 CIDEr across six benchmarks and aligning with human-interpretable relevance.

This study introduces EnterpriseMem-Bench, a multi-turn Text-to-SQL benchmark with 300 sessions and 1,400 turns. Evaluating five frontier models reveals: stateless models collapse to zero accuracy by Turn 3; memory complexity does not monotonically improve performance, with working memory dominating; Claude Sonnet 4.6 shows generational regression on SEC EDGAR; and under reasoning, Claude error distributions become mono-modal.

This paper proposes a generalizable framework for cultural evaluation and intervention in LLMs, using scenario-based behavioral probing and activation steering to shift internal alignments without retraining. Experiments reveal latent entanglement between cultural dimensions, limiting precise value alignment.