SCAN-Planner is a spatial collision-aware local planning framework for quadruped robots, using a yaw-aware twin-cylinder footprint model and projected A* search to navigate narrow passages, cluttered indoor scenes, and large-scale 3D unstructured environments safely and efficiently.

This research proposes a novel multi-layered mathematical model using category theory and sheaf theory to describe the Software Component Ensemble Language (SCEL) for autonomous agent systems. The model treats components as points, ensembles as open sets, and distributed knowledge as sheaf data. Information sharing is modeled as 'gluing' local data, and system failures are quantified as topological obstructions via sheaf cohomology. This approach transforms verification of complex distributed systems into geometric analysis, providing structural insights for robust autonomic system design.

This paper presents an invariant extended Kalman filtering (InEKF) approach for real-time state estimation of humanoid robots on non-inertial ground using only onboard proprioceptive sensing. It estimates the robot's base position and velocity relative to the moving ground without direct ground motion measurements, using foot-mounted IMUs and kinematic constraints. Experiments on Digit robot show 96% faster convergence and 80% lower position errors on swaying ground, and average errors under 9 cm on rotating ground.

This paper introduces Playful Agentic Robot Learning, where an embodied coding agent uses self-directed play as a continual skill-learning stage. The proposed RATs framework proposes exploratory tasks, executes code policies, verifies progress, and distills successful executions into a persistent skill library. Experiments show significant improvements on LIBERO-PRO and MolmoSpaces, and transferable skills boost performance without fine-tuning.

The paper presents DiffusionVS, a diffusion-based visual servoing method that uses conditional denoising to generate camera velocity and online training for improved generalization. It achieves nearly 100% success in simulation and 93% in physical experiments, and can be integrated into existing visual servoing networks to boost performance.

3D Scene Graphs (3DSGs) have emerged as a powerful representation for spatial AI, combining geometric grounding with semantic and relational abstractions. They are relevant to robotics and computer vision problems including manipulation, navigation, task planning, and scene understanding. However, the field is fragmented across different communities with distinct formulations, construction pipelines, and evaluation protocols, making it difficult to compare methods and assess challenges for real-world deployment. This survey provides a unified and critical review of 3DSGs, focusing on open challenges and future directions. It formalizes 3DSGs under a common definition, analyzes modeling choices (node/edge attributes, hierarchical structure, dynamic scenes, affordance-aware extensions), rev…

WorkBenchMark is a LEGO Duplo-based robotic assembly benchmark inspired by the RoboCup Smart Manufacturing League. It provides 400 tasks across four complexity tiers and introduces an open-vocabulary Assembly-by-Disassembly baseline. The planning-based pipeline outperforms modern vision-language-action approaches on all tiers. The benchmark, simulation environment, and baseline implementation will be released openly.

Researchers built a robot called Robotroller that actuates an Atari CX40+ controller and a device called Atari Devbox that renders game frames and reward signals from the Arcade Learning Environment. Together with an off-the-shelf camera and desktop computer, the system forms Physical Atari, a robust (bearings for movement, high-frequency servo monitoring) and accessible (under $1,000, 3D-printed parts) platform for studying real-world reinforcement learning. Weeks of non-stop experiments validated that RL algorithms can learn directly on robots, while showing that small distribution shifts between training and deployment severely degrade policy performance, emphasizing the need for on-device adaptation.

Researchers propose TeleMorpher, a one-shot framework for simultaneous motion and location editing in videos using diffusion models. It disentangles protagonist and background, uses pose warping with motion priors, and introduces new evaluation metrics. Experiments show superior performance on in-the-wild videos and the TaiChi dataset.

The paper proposes learning an asynchronous schedule for denoising in multi-representation latent diffusion models. It introduces a schedule-corrected objective and a flexible parametric class that is convex and monotone. The schedule is learned with minimal additional compute (<1%). On ImageNet 256x256, the method achieves FID 1.05 in 200 epochs (matching a 800-epoch baseline) and FID 1.02 in 600 epochs (outperforming a 1B-parameter model). Unguided results also show significant improvements.

GB-LSR introduces a fixed-grid local spectral representation for continuous image reconstruction using a single shared global bandwidth. It outperforms existing methods like LIIF, LTE, and WIRE on native reconstruction benchmarks and offers significant speedups in super-resolution tasks.

Language-Instructed Vision Embeddings (LIVE) uses language as high-level guidance to produce task-centric embeddings at inference time, removing the need for task-specific retraining. It reduces visual hallucinations by 34 points on MMVP, surpasses much larger vision-language models on VQA, and generalizes to unseen instructions and tasks.

Mix-QVLA is a task-evidence-aware mixed-precision post-training quantization framework for VLA models. It evaluates whether quantization preserves task-relevant evidence and dynamically adjusts layer precision, achieving high accuracy with significant memory reduction and speedup. On LIBERO, it reduces memory from 15.4 GB to 4.1 GB, retains 96.3% success rate, and achieves 1.52x inference speedup.

This paper proposes PerceptionDLM, a multimodal diffusion language model optimized for efficient parallel region perception. Leveraging the parallel decoding nature of diffusion language models, it introduces efficient prompting and structured attention masking to enable simultaneous perception of multiple masked regions, significantly improving inference efficiency. A new benchmark, ParaDLC-Bench, is constructed to evaluate parallelism in visual perception. Experiments show competitive performance with substantial speed improvements for multi-region tasks.

A new approach called ImageWAM replaces video generation with image editing for robot action prediction, achieving superior performance and significantly lower computational cost.

LooseControlVideo is a novel framework that enables intuitive and expressive text-to-video generation control via sparse, oriented 3D boxes as a "blocking" proxy. It fine-tunes a Wan 2.2 backbone with DNOCS encoding for 3D size, orientation, and depth-ordered occlusions, supporting localized refinements without global disruption. Evaluations on nuScenes, HO-3D, and BEHAVE benchmarks show significant improvements: 1.2x-3x in Trajectory Error, 2x in Rigid Motion Consistency, and 1.5x-2x in Occlusion Accuracy over existing methods.

LEAP introduces a training curriculum for Vision Transformer feature-based knowledge distillation that uses the teacher’s intermediate feature maps as progressively harder targets, enabling the student to build a foundation before tackling higher-level abstractions. It achieves significant accuracy gains and training efficiency improvements across multiple benchmarks.

Researchers introduce the first billion-parameter generative foundation model for chest radiograph synthesis, with over 1.3B parameters trained on 1.2M radiographs and expert metadata. The model supports controllable generation across demographics, views, and pathologies, achieving near-clinical indistinguishability.

This paper introduces self-function vectors to directly estimate aleatoric uncertainty in in-context learning under a Bayesian framework, along with the first rigorous evaluation protocol to separate aleatoric from epistemic uncertainty. Experiments show the method reliably measures LLM prediction uncertainty and can be used for hallucination detection. Accepted to ACL 2026.

A new comprehensive survey indexes 120 sign-language datasets across 35 sign languages, identifying key challenges such as modality imbalance and annotation inconsistency. The authors introduce a standardized 24-field datasheet and a public GitHub repository to promote reproducible research in sign-language technologies.

Proposes LUCID, a hallucination detection method that leverages LLM attention scores, KG semantics, and structural information via graph neural networks, achieving state-of-the-art performance on nine datasets.

Researchers propose Causal Attribution Pruning (CAP), a training-free method that identifies critical attention heads in LLMs by measuring their causal impact on reasoning tasks. CAP achieves up to 61% relative accuracy improvement over Wanda at 20% sparsity on ARC-Challenge, and consistently outperforms at moderate sparsity levels.

This paper reveals that query position critically affects in-context learning in diffusion LLMs due to a recency effect in attention. The authors propose Average Confidence metric and Auto-ICL, a training-free strategy to optimize query placement, achieving robust performance across tasks.

A new study introduces an error taxonomy for LLMs in hardware design, revealing a ceiling on VerilogEval at 90.8% pass rate due to unsolvable functional errors. Alignment techniques only teach compilation, not reasoning.

This study fine-tunes seven large language models on Arabic and evaluates zero-shot reading comprehension on Semitic and non-Semitic languages. No evidence of Semitic-specific transfer was found; improvements are due to task-format alignment, not linguistic knowledge.

This study investigates using Google's Gemini and Gemma LLMs to automate detection of EQ-5D studies in PubMed from abstracts. A multi-phase framework with few-shot prompting, weight ensembling, and soft stacking meta-classifier is proposed. Evaluating nine LLMs on expert-labeled data, the weighted ensemble of gemini-2.5-pro, gemma-3-12b, and gemma-3-27b achieved 0.74 weighted F1-score and accuracy, outperforming individual models. Ensembling improved precision-recall balance, and soft stacking enhanced reliability and interpretability. Results suggest ensemble LLMs are a reliable, scalable approach for automating biomedical screening.