三菱电机

Industrial robotic manipulators enhance manufacturing productivity and safety, but can experience anomalies that disrupt production and compromise safety. The complexity and high variability of modern robotics applications make anomaly detection (AD) particularly challenging. Data-driven semi-supervised approaches are emerging as a promising solution for this task. To overcome the lack of sufficiently rich labeled datasets, models are trained using only normal data, and anomalies are identified by detecting deviations from the normal distribution. To ensure reliable deployment of this technology, it is crucial to design models that are robust against variations in working conditions. At the same time, the availability of public datasets with data from actual systems is fundamental to compare different strategies. In this paper, we introduce Mitsubishi Electric Anomaly Detection (ME-AD), a novel dataset with data capturing real-world, progressively evolving mechanical faults in robotic manipulators during pick-and-place operations. Unlike existing datasets, ME-AD provides authentic sensor data documenting gradual joint degradation up to a mechanical fault during a complete run-to-failure. Furthermore, we propose a Gaussian Process Regression (GPR)-based anomaly detection method that models robot dynamics and employs a novel uncertainty-aware anomaly score to detect anomalies, offering generalization across unseen motion conditions. Our extensive evaluations demonstrate strong anomaly-detection performance and cross-trajectory transfer to out-of-distribution motion profiles and execution speeds.

We present an optimization-based approach for fuel-efficient spacecraft rendezvous to Gateway, a lunar space station architecture based on a near rectilinear halo orbit (NRHO) around the Earth-Moon L2 point. Our main contribution is a method that ensures safety at all times in the presence of Gaussian uncertainty while incorporating disturbance compensation feedback through correction maneuvers. Specifically, the approach: i) ensures passive safety and satisfies path constraints at all times, ii) satisfies the specifications for critical decision points along the trajectory, and iii) accounts for common uncertainties arising from orbital insertion maneuvers, actuation, and navigation measurements. To account for these uncertainties, we use both chance constraints and disturbance compensation feedback to bound the effects of errors. We leverage sequential convex programming (SCP) together with an isoperimetric reformulation of continuous-time path constraints, including passive safety, to eliminate the risk of inter-sample constraint violations common in existing methods. This enables sparse time discretization that supports efficient offline mission design and analysis and aligns with operational preferences for less frequent control actions. We demonstrate the method on a case study representative of Gateway rendezvous operations.

Real-world scenes often feature multiple humans interacting with multiple objects in ways that are causal, goal-oriented, or cooperative. Yet existing 3D human-object interaction (HOI) benchmarks consider only a fraction of these complex interactions. To close this gap, we present MMHOI - a large-scale, Multi-human Multi-object Interaction dataset consisting of images from 12 everyday scenarios. MMHOI offers complete 3D shape and pose annotations for every person and object, along with labels for 78 action categories and 14 interaction-specific body parts, providing a comprehensive testbed for next-generation HOI research. Building on MMHOI, we present MMHOI-Net, an end-to-end transformer-based neural network for jointly estimating human-object 3D geometries, their interactions, and associated actions. A key innovation in our framework is a structured dual-patch representation for modeling objects and their interactions, combined with action recognition to enhance the interaction prediction. Experiments on MMHOI and the recently proposed CORE4D datasets demonstrate that our approach achieves state-of-the-art performance in multi-HOI modeling, excelling in both accuracy and reconstruction quality.

Neural audio codecs (NACs), which use neural networks to generate compact audio representations, have garnered interest for their applicability to many downstream tasks, especially quantized codecs due to their compatibility with large language models. However, unlike text, speech conveys not only linguistic content but also rich paralinguistic features. Encoding these elements in an entangled fashion may be suboptimal, as it limits flexibility. For instance, voice conversion (VC) aims to convert speaker characteristics while preserving the original linguistic content, which requires a disentangled representation. Inspired by VC methods utilizing k-means quantization with self-supervised features to disentangle phonetic information, we develop a discrete NAC capable of structured disentanglement. Experimental evaluations show that our approach achieves reconstruction performance on par with conventional NACs that do not explicitly perform disentanglement, while also matching the effectiveness of conventional VC techniques.

Recent advances in Large Language Models (LLMs) have improved multi-step reasoning. Most approaches rely on Chain-of-Thought (CoT) rationales. Previous studies have shown that LLMs often generate logically inconsistent reasoning steps even when their final answers are correct. These inconsistencies reduce the reliability of step-level reasoning. We propose GeoSteer, a manifold-based framework that improves the quality of intermediate reasoning. The method consists of: (1) constructing a CoT dataset with segment-level scores, (2) training a Variational Autoencoder (VAE) model and a quality estimation model to learn a low-dimensional manifold of high-quality CoT trajectories, and (3) steering hidden states of target LLMs toward higher-quality regions in the latent space. This update in a latent space behaves like a natural-gradient adjustment in the original hidden-state space. It ensures geometrically coherent steering. We evaluate GeoSteer on the GSM8k dataset using the Qwen3 series. We measure via answer accuracy and overall reasoning performance. GeoSteer improved the exact match accuracy by up to 2.6 points. It also enhanced the pairwise win rate by 5.3 points. These results indicate that GeoSteer provides an effective and controllable mechanism for improving the quality of intermediate reasoning in LLMs.