Recent advances in vision language models (VLMs) have demonstrated remarkable potential in embodied navigation tasks. However, existing robot-centric datasets primarily focus on traditional 3D tasks such as perception and prediction, lacking adequate support for vision-language tasks. Vision-language-navigation (VLN) is a key capability for achieving human-like and interpretable navigation in complex environments. In this study, we present CoT-VLNBench, the first large-scale benchmark and dataset designed for chain-of-thought (CoT) reasoning in quadruped robot navigation. Our dataset encompasses a diverse range of indoor and outdoor scenes, multi-step navigation trajectories, and rich natural language instructions, all annotated with fine-grained CoT reasoning traces. Specifically, it contains 175K frames, 5.25M 3D bounding boxes, and 875K vision–question–answer (VQA) pairs. This comprehensive resource enables thorough evaluation of embodied agents’ perceptual and step-by-step reasoning abilities. Furthermore, we propose a novel CoT-VLN model, a state-of-the-art 7B VLN model that integrates visual, linguistic, and reasoning modules, to facilitate interpretable and effective navigation. Extensive experiments demonstrate that our approach significantly outperforms existing non-VLMs baselines on the new benchmark, underscoring the importance of CoT-VLN in embodied navigation. We hope that CoT-VLNBench will serve as a valuable resource to advance research at the intersection of robotics, vision, language, and reasoning.

Large language models (LLMs) have progressed rapidly in complex reasoning and question answering, yet LLM hallucination remains a central bottleneck that hinders practical deployment, especially for commercial black-box LLMs accessible only via APIs.Existing uncertainty quantification methods typically depend on computationally expensive multiple sampling or internal parameters, which prevents real-time estimation and fails to capture information implicit in the blackbox reasoning process.To address this issue, we propose Distribution-Aligned Adversarial Distillation (DisAAD), which introduces a generation-discrimination architecture to guide a lightweight proxy model to learn the highquality regions of the output distribution of the black-box LLM, thus effectively endowing it with the ability to "know whether the blackbox LLM knows or not".Subsequently, we use the proxy model to reproduce the specific responses of the black-box LLM and estimate the corresponding uncertainty based on evidence learning.Extensive experiments have verified the effectiveness and promise of our proposed method, indicating that a proxy model even one that only accounts for 1% of the target LLM's size can achieve reliable uncertainty quantification.Our model and related resources are released at https://github.com/huizi-Cui/ DisAAD.

The trade-off between system performance and energy efficiency (service time) is critical for battery-based embedded systems. Most of the previous work focuses on saving energy in a deterministic way by taking the average or worst scenario into account. However, such deterministic approaches usually are inappropriate in modeling energy consumption because of uncertainties in conditional instructions on processors and time-varying external environments (e.g., fluctuant network bandwidth and different user inputs). By adopting a probabilistic approach, this paper proposes a model and a set of algorithms to address the Processor and Voltage Assignment with Probability (PVAP) problem of data-dependent aperiodic tasks in real-time embedded systems, ensuring that all the tasks can be done under the time constraint with a guaranteed probability. We adopt a task DAG (Directed Acyclic Graph) to model the PVAP problem. We first use a processor scheduling algorithm to map the task DAG onto a set of voltage-variable processors, and then use our dynamic programming algorithm to assign a proper voltage to each task. Finally, to escape from local optima, a local search with restarts searches the optimal solution from candidate solutions by updating the objective function, until the stop criteria are reached or a time bound is elapsed. The experimental results demonstrate that for probability 1.0, our approach yields slightly better results than the well-known algorithms like ASAP/ALAP (As Soon As Possible/As Late As Possible) and ILP (Integer Linear Programming) with/without DVS (Dynamic Voltage Scaling). However, for probabilities 0.8 and 0.9, our approach significantly outperforms those algorithms (maximum improvement of 50.3 percent).

Indoor localization becomes increasingly important as context-aware applications gain popularity in mobile users. A promising approach for indoor localization is to leverage the pervasive WiFi infrastructure via fingerprinting-based inference. However, a WiFi device must frequently scan for WiFi signals during localization, leading to high power consumption. Moreover, switching to the scanning mode introduces inevitable disruptions to data communication of WiFi interface. This paper presents a new indoor localization system called ZiFind that exploits the cross-technology interference in the unlicensed 2.4 GHz frequency spectrum. ZiFind utilizes low-power ZigBee interface to collect WiFi interference signals and adopts digital signal processing techniques to extract unique signatures as fingerprints for localization. To deal with the noise in the fingerprints, we design a new learning algorithm called R-KNN that can improve the accuracy of localization by assigning different weights to fingerprint features according to their importance. We implement ZiFind on TelosB motes and evaluate its performance through extensive experiments in a 16,000 ft2office building floor consisting of 28 rooms. Our results show that ZiFind leads to significant power saving compared with existing approaches based on WiFi interface, and yields satisfactory localization accuracy in a range of realistic settings.