Multimodal large language models (MLLMs) show promise in tasks like visual question answering (VQA) but still face challenges in multimodal reasoning. Recent works adapt agentic frameworks or chain-of-thought (CoT) reasoning to improve performance. However, CoT-based multimodal reasoning often demands costly data annotation and fine-tuning, while agentic approaches relying on external tools risk introducing unreliable output from these tools. In this paper, we propose Seeing and Reasoning with Confidence (SRICE), a training-free multimodal reasoning framework that integrates external vision models with uncertainty quantification (UQ) into an MLLM to address these challenges. Specifically, SRICE guides the inference process by allowing MLLM to autonomously select regions of interest through multi-stage interactions with the help of external tools. We propose to use a conformal prediction-based approach to calibrate the output of external tools and select the optimal tool by estimating the uncertainty of an MLLM's output. Our experiment shows that the average improvement of SRICE over the base MLLM is 4.6% on five datasets and the performance on some datasets even outperforms fine-tuning-based methods, revealing the significance of ensuring reliable tool use in an MLLM agent.

The increasing availability of traffic videos functioning on a 24/7/365 time scale has the great potential of increasing the spatio-temporal coverage of traffic accidents, which will help improve traffic safety. However, analyzing footage from hundreds, if not thousands, of traffic cameras in a 24/7/365 working protocol remains an extremely challenging task, as current vision-based approaches primarily focus on extracting raw information, such as vehicle trajectories or individual object detection, but require laborious post-processing to derive actionable insights. We propose SeeUnsafe, a new framework that integrates Multimodal Large Language Model (MLLM) agents to transform video-based traffic accident analysis from a traditional extraction-then-explanation workflow to a more interactive, conversational approach. This shift significantly enhances processing throughput by automating complex tasks like video classification and visual grounding, while improving adaptability by enabling seamless adjustments to diverse traffic scenarios and user-defined queries. Our framework employs a severity-based aggregation strategy to handle videos of various lengths and a novel multimodal prompt to generate structured responses for review and evaluation and enable fine-grained visual grounding. We introduce IMS (Information Matching Score), a new MLLM-based metric for aligning structured responses with ground truth. We conduct extensive experiments on the Toyota Woven Traffic Safety dataset, demonstrating that SeeUnsafe effectively performs accident-aware video classification and visual grounding by leveraging off-the-shelf MLLMs. Source code will be available at \url{https://github.com/ai4ce/SeeUnsafe}.

Photorealistic simulators are essential for the training and evaluation of vision-centric autonomous vehicles (AVs). At their core is Novel View Synthesis (NVS), a crucial capability that generates diverse unseen viewpoints to accommodate the broad and continuous pose distribution of AVs. Recent advances in radiance fields, such as 3D Gaussian Splatting, achieve photorealistic rendering at real-time speeds and have been widely used in modeling large-scale driving scenes. However, their performance is commonly evaluated using an interpolated setup with highly correlated training and test views. In contrast, extrapolation, where test views largely deviate from training views, remains underexplored, limiting progress in generalizable simulation technology. To address this gap, we leverage publicly available AV datasets with multiple traversals, multiple vehicles, and multiple cameras to build the first Extrapolated Urban View Synthesis (EUVS) benchmark. Meanwhile, we conduct quantitative and qualitative evaluations of state-of-the-art Gaussian Splatting methods across different difficulty levels. Our results show that Gaussian Splatting is prone to overfitting to training views. Besides, incorporating diffusion priors and improving geometry cannot fundamentally improve NVS under large view changes, highlighting the need for more robust approaches and large-scale training. We have released our data to help advance self-driving and urban robotics simulation technology.

Navigating dynamic urban environments presents significant challenges for embodied agents, requiring advanced spatial reasoning and adherence to common-sense norms. Despite progress, existing visual navigation methods struggle in map-free or off-street settings, limiting the deployment of autonomous agents like last-mile delivery robots. To overcome these obstacles, we propose a scalable, data-driven approach for human-like urban navigation by training agents on thousands of hours of in-the-wild city walking and driving videos sourced from the web. We introduce a simple and scalable data processing pipeline that extracts action supervision from these videos, enabling large-scale imitation learning without costly annotations. Our model learns sophisticated navigation policies to handle diverse challenges and critical scenarios. Experimental results show that training on large-scale, diverse datasets significantly enhances navigation performance, surpassing current methods. This work shows the potential of using abundant online video data to develop robust navigation policies for embodied agents in dynamic urban settings. Project homepage is at https://ai4ce.github.io/CityWalker/.

Visual localization, which estimates a camera's pose within a known scene, is a long-standing challenge in vision and robotics. Recent end-to-end methods that directly regress camera poses from query images have gained attention for fast inference. However, existing methods often struggle to generalize to unseen views. In this work, we aim to unleash the power of data synthesis to promote the generalizability of pose regression. Specifically, we lift real 2D images into 3D Gaussian Splats with varying appearance and deblurring abilities, which are then used as a data engine to synthesize more posed images. To fully leverage the synthetic data, we build a two-branch joint training pipeline, with an adversarial discriminator to bridge the syn-to-real gap. Experiments on established benchmarks show that our method outperforms state-of-the-art end-to-end approaches, reducing translation and rotation errors by 50% and 21.6% on indoor datasets, and 35.56% and 38.7% on outdoor datasets. We also validate the effectiveness of our method in dynamic driving scenarios under varying weather conditions. Notably, as data synthesis scales up, our method exhibits a growing ability to interpolate and extrapolate training data for localizing unseen views. Project Page: https://ai4ce.github.io/RAP/

Continuously adapting pre-trained models to local data on resource constrained edge devices is the $\emph{last mile}$ for model deployment. However, as models increase in size and depth, backpropagation requires a large amount of memory, which becomes prohibitive for edge devices. In addition, most existing low power neural processing engines (e.g., NPUs, DSPs, MCUs, etc.) are designed as fixed-point inference accelerators, without training capabilities. Forward gradients, solely based on directional derivatives computed from two forward calls, have been recently used for model training, with substantial savings in computation and memory. However, the performance of quantized training with fixed-point forward gradients remains unclear. In this paper, we investigate the feasibility of on-device training using fixed-point forward gradients, by conducting comprehensive experiments across a variety of deep learning benchmark tasks in both vision and audio domains. We propose a series of algorithm enhancements that further reduce the memory footprint, and the accuracy gap compared to backpropagation. An empirical study on how training with forward gradients navigates in the loss landscape is further explored. Our results demonstrate that on the last mile of model customization on edge devices, training with fixed-point forward gradients is a feasible and practical approach.

Recent advancements in brain-computer interfaces (BCIs) have enabled the decoding of lexical tones from intracranial recordings, offering the potential to restore the communication abilities of speech-impaired tonal language speakers. However, data heterogeneity induced by both physiological and instrumental factors poses a significant challenge for unified invasive brain tone decoding. Traditional subject-specific models, which operate under a heterogeneous decoding paradigm, fail to capture generalized neural representations and cannot effectively leverage data across subjects. To address these limitations, we introduce Homogeneity-Heterogeneity Disentangled Learning for neural Representations (H2DiLR), a novel framework that disentangles and learns both the homogeneity and heterogeneity from intracranial recordings across multiple subjects. To evaluate H2DiLR, we collected stereoelectroencephalography (sEEG) data from multiple participants reading Mandarin materials comprising 407 syllables, representing nearly all Mandarin characters. Extensive experiments demonstrate that H2DiLR, as a unified decoding paradigm, significantly outperforms the conventional heterogeneous decoding approach. Furthermore, we empirically confirm that H2DiLR effectively captures both homogeneity and heterogeneity during neural representation learning. The human language system, with its intricate and expansive syntactic structure, enables rich and complex communication. Decoding spoken language from within human brains has emerged as a significant topic of interest in neuroscience (Anumanchipalli et al., 2019; Willett et al., 2023; Feng et al., 2023; Lu et al., 2023; Liu et al., 2023). The decoding of vocal tone from brain measurements (Lu et al., 2023; Liu et al., 2023) is of particular research interest, due to the prominence of tonal languages, which make up over 60% of the world's languages (Yip, 2002) and are spoken by approximately one-third of the global population (Dryer & Haspelmath, 2013). In these languages, tone plays a critical role in distinguishing lexical meaning at the syllable level. Mandarin, for instance, is a widely spoken tonal language that has an extensive inventory of over 50,000 characters, with each associated with a syllable composed of an initial, a final, and a tone (Duanmu, 2007).

Vision Language Models (VLMs) have recently been adopted in robotics for their capability in common sense reasoning and generalizability. Existing work has applied VLMs to generate task and motion planning from natural language instructions and simulate training data for robot learning. In this work, we explore using VLM to interpret human demonstration videos and generate robot task planning. Our method integrates keyframe selection, visual perception, and VLM reasoning into a pipeline. We named it SeeDo because it enables the VLM to ''see'' human demonstrations and explain the corresponding plans to the robot for it to ''do''. To validate our approach, we collected a set of long-horizon human videos demonstrating pick-and-place tasks in three diverse categories and designed a set of metrics to comprehensively benchmark SeeDo against several baselines, including state-of-the-art video-input VLMs. The experiments demonstrate SeeDo's superior performance. We further deployed the generated task plans in both a simulation environment and on a real robot arm.

Humans effortlessly integrate common-sense knowledge with sensory input from vision and touch to understand their surroundings. Emulating this capability, we introduce FusionSense, a novel 3D reconstruction framework that enables robots to fuse priors from foundation models with highly sparse observations from vision and tactile sensors. FusionSense addresses three key challenges: (i) How can robots efficiently acquire robust global shape information about the surrounding scene and objects? (ii) How can robots strategically select touch points on the object using geometric and common-sense priors? (iii) How can partial observations such as tactile signals improve the overall representation of the object? Our framework employs 3D Gaussian Splatting as a core representation and incorporates a hierarchical optimization strategy involving global structure construction, object visual hull pruning and local geometric constraints. This advancement results in fast and robust perception in environments with traditionally challenging objects that are transparent, reflective, or dark, enabling more downstream manipulation or navigation tasks. Experiments on real-world data suggest that our framework outperforms previously state-of-the-art sparse-view methods. All code and data are open-sourced on the project website.

However, their performance in place recognition is still underexplored. In this work, we introduce multimodal LLMs (MLLMs) to visual place recognition (VPR), where a robot must localize itself using visual observations. Our key design is to use vision-based retrieval to propose several candidates and then leverage language-based reasoning to carefully inspect each candidate for a final decision. Specifically, we leverage the robust visual features produced by off-the-shelf vision foundation models (VFMs) to obtain several candidate locations. We then prompt an MLLM to describe the differences between the current observation and each candidate in a pairwise manner, and reason about the best candidate based on these descriptions. Our results on three datasets demonstrate that integrating the general-purpose visual features from VFMs with the reasoning capabilities of MLLMs already provides an effective place recognition solution, without any VPR-specific supervised training. We believe our work can inspire new possibilities for applying and designing foundation models, i.e., VFMs, LLMs, and MLLMs, to enhance the localization and navigation of mobile robots.