20 August 2025 in IEEE Robotics and Automation Letters
Robust grasping in cluttered environments remains an open challenge in robotics. While benchmark datasets have significantly advanced deep learning methods, they mainly focus on simplistic scenes with light occlusion and insufficient diversity, limiting their applicability to practical scenarios. We present GraspClutter6D, a large-scale real-world grasping dataset featuring: (1) 1,000 highly cluttered scenes with dense arrangements (14.1 objects/scene, 62.6% occlusion), (2) comprehensive coverage across 200 objects in 75 environment configurations (bins, shelves, and tables) captured using four RGB-D cameras from multiple viewpoints, and (3) rich annotations including 736 K 6D object poses and 9.3B feasible robotic grasps for 52 K RGB-D images. We benchmark state-of-the-art segmentation, object pose estimation, and grasp detection methods to provide key insights into challenges in cluttered environments. Additionally, we validate the dataset's effectiveness as a training resource, demonstrating that grasping networks trained on GraspClutter6D significantly outperform those trained on existing datasets in both simulation and real-world experiments.
16 April 2026 in IEEE Access
Maintenance and operation of critical energy infrastructure require extreme precision and strict adherence to safety protocols to prevent catastrophic failures. While recent advancements in general-purpose Vision-Language-Action (VLA) foundation models have shown promise in robotics, their inherent stochasticity and lack of procedural precision often result in unacceptable safety risks in high-stakes industrial environments. To address these limitations, this paper proposes a novel Hierarchical Multi-Robot Orchestration Framework designed for the coordinated control of heterogeneous robots in energy facility maintenance. The proposed framework decouples high-level cognitive reasoning from low-level execution by utilizing a local Large Language Model (LLM) as a strategic orchestrator. Grounded in a Structured Environmental Representation (M), the LLM employs Chain-of-Thought (CoT) reasoning to decompose complex maintenance missions into verifiable sub-tasks, effectively bridging the gap between linguistic intent and physical constraints. These tasks are then dispatched to specialized execution modules: a manipulation unit utilizing Action Chunking with Transformers (ACT) adapted for high-precision industrial tasks, an autonomous navigation unit for Simultaneous Localization and Mapping (SLAM)-based pathfinding, and a vision-guided logistics unit for retrieval. Experimental evaluations in a simulated power plant environment validate the efficacy of the framework. A comparative ablation study demonstrates that utilizing structured environmental grounding significantly enhances the reasoning reliability of large-scale models compared to unstructured text baselines. Specifically, the GPT-OSS (20B) model achieved a peak planning success rate of 91.7%, with a notable proficiency in handling long-horizon Strategic commands (73.3%), validating that structural clarity is essential for complex causal inference. Furthermore, the integrated execution layer demonstrated exceptional reliability, achieving a 90% success rate specifically in distribution panel maintenance tasks, confirming that the hierarchical decoupling of probabilistic reasoning and deterministic execution provides a reliable solution for autonomous maintenance.
21 May 2026 in IEEE Access
Instance segmentation is a crucial task in computer vision with numerous applications. Recent approaches incorporating refinement modules have shown promise. However, their reliance on indirect guidance often results in suboptimal performance and high computational costs.We introduce the self-error estimation and dual refinement (SEED), a simple yet powerful plug-in module that significantly enhances instance segmentation quality through explicit fine-grained guidance. SEED identifies proposed self-error, defined as pixel-wise discrepancies between initial predictions and ground truth. It then leverages this information to guide the dual refinement of bounding boxes and masks. SEED seamlessly integrates into various existing methods by being added as a plug-in, and enables targeted and effective improvements in both detection and segmentation quality. Extensive experiments demonstrate that SEED significantly improves the performance of state-of-the-art models across diverse datasets, including natural, driving, and robotic scenes while maintaining computational efficiency, outperforming existing refinement-based high-quality instance segmentation methods. The code is available at https://github.com/gist-ailab/seed.