Computer use agents are LLM-based agents that can directly interact with a graphical user interface, by processing screenshots or accessibility trees. While these systems are gaining popularity, their safety has been largely overlooked, despite the fact that evaluating and understanding their potential for harmful behavior is essential for widespread adoption. To address this gap, we introduce OS-Harm, a new benchmark for measuring safety of computer use agents. OS-Harm is built on top of the OSWorld environment and aims to test models across three categories of harm: deliberate user misuse, prompt injection attacks, and model misbehavior. To cover these cases, we create 150 tasks that span several types of safety violations (harassment, copyright infringement, disinformation, data exfiltration, etc.) and require the agent to interact with a variety of OS applications (email client, code editor, browser, etc.). Moreover, we propose an automated judge to evaluate both accuracy and safety of agents that achieves high agreement with human annotations (0.76 and 0.79 F1 score). We evaluate computer use agents based on a range of frontier models - such as o4-mini, Claude 3.7 Sonnet, Gemini 2.5 Pro - and provide insights into their safety. In particular, all models tend to directly comply with many deliberate misuse queries, are relatively vulnerable to static prompt injections, and occasionally perform unsafe actions. The OS-Harm benchmark is available at https://github.com/tml-epfl/os-harm.
We present WorldVLA, an autoregressive action world model that unifies action and image understanding and generation. Our WorldVLA intergrates Vision-Language-Action (VLA) model and world model in one single framework. The world model predicts future images by leveraging both action and image understanding, with the purpose of learning the underlying physics of the environment to improve action generation. Meanwhile, the action model generates the subsequent actions based on image observations, aiding in visual understanding and in turn helps visual generation of the world model. We demonstrate that WorldVLA outperforms standalone action and world models, highlighting the mutual enhancement between the world model and the action model. In addition, we find that the performance of the action model deteriorates when generating sequences of actions in an autoregressive manner. This phenomenon can be attributed to the model's limited generalization capability for action prediction, leading to the propagation of errors from earlier actions to subsequent ones. To address this issue, we propose an attention mask strategy that selectively masks prior actions during the generation of the current action, which shows significant performance improvement in the action chunk generation task.
This paper presents ESFP, an end-to-end pipeline that converts monocular RGB video into executable joint trajectories for a low-cost 4-DoF desktop arm. ESFP comprises four sequential modules. (1) Estimating: ROMP lifts each frame to a 24-joint 3-D skeleton. (2) Smoothing: the proposed HPSTM-a sequence-to-sequence Transformer with self-attention-combines long-range temporal context with a differentiable forward-kinematics decoder, enforcing constant bone lengths and anatomical plausibility while jointly predicting joint means and full covariances. (3) Filtering: root-normalized trajectories are variance-weighted according to HPSTM's uncertainty estimates, suppressing residual noise. (4) Pose-Mapping: a geometric retargeting layer transforms shoulder-elbow-wrist triples into the uArm's polar workspace, preserving wrist orientation.
Panoramic image processing is essential for omni-context perception, yet faces constraints like distortions, perspective occlusions, and limited annotations. Previous unsupervised domain adaptation methods transfer knowledge from labeled pinhole data to unlabeled panoramic images, but they require access to source pinhole data. To address these, we introduce a more practical task, i.e., Source-Free Occlusion-Aware Seamless Segmentation (SFOASS), and propose its first solution, called UNconstrained Learning Omni-Context Knowledge (UNLOCK). Specifically, UNLOCK includes two key modules: Omni Pseudo-Labeling Learning and Amodal-Driven Context Learning. While adapting without relying on source data or target labels, this framework enhances models to achieve segmentation with 360{\deg} viewpoint coverage and occlusion-aware reasoning. Furthermore, we benchmark the proposed SFOASS task through both real-to-real and synthetic-to-real adaptation settings. Experimental results show that our source-free method achieves performance comparable to source-dependent methods, yielding state-of-the-art scores of 10.9 in mAAP and 11.6 in mAP, along with an absolute improvement of +4.3 in mAPQ over the source-only method. All data and code will be made publicly available at https://github.com/yihong-97/UNLOCK.
3D Semantic Occupancy Prediction is crucial for autonomous driving, providing a dense, semantically rich environmental representation. However, existing methods focus on in-distribution scenes, making them susceptible to Out-of-Distribution (OoD) objects and long-tail distributions, which increases the risk of undetected anomalies and misinterpretations, posing safety hazards. To address these challenges, we introduce Out-of-Distribution Semantic Occupancy Prediction, targeting OoD detection in 3D voxel space. To fill the gaps in the dataset, we propose a Synthetic Anomaly Integration Pipeline that injects synthetic anomalies while preserving realistic spatial and occlusion patterns, enabling the creation of two datasets: VAA-KITTI and VAA-KITTI-360. We introduce OccOoD, a novel framework integrating OoD detection into 3D semantic occupancy prediction, with Voxel-BEV Progressive Fusion (VBPF) leveraging an RWKV-based branch to enhance OoD detection via geometry-semantic fusion. Experimental results demonstrate that OccOoD achieves state-of-the-art OoD detection with an AuROC of 67.34% and an AuPRCr of 29.21% within a 1.2m region, while maintaining competitive occupancy prediction performance. The established datasets and source code will be made publicly available at https://github.com/7uHeng/OccOoD.
Vision-language-action (VLA) models extend vision-language models (VLM) by integrating action generation modules for robotic manipulation. Leveraging strengths of VLM in vision perception and instruction understanding, VLA models exhibit promising generalization across diverse manipulation tasks. However, applications demanding high precision and accuracy reveal performance gaps without further adaptation. Evidence from multiple domains highlights the critical role of post-training to align foundational models with downstream applications, spurring extensive research on post-training VLA models. VLA model post-training aims to address the challenge of improving an embodiment's ability to interact with the environment for the given tasks, analogous to the process of humans motor skills acquisition. Accordingly, this paper reviews post-training strategies for VLA models through the lens of human motor learning, focusing on three dimensions: environments, embodiments, and tasks. A structured taxonomy is introduced aligned with human learning mechanisms: (1) enhancing environmental perception, (2) improving embodiment awareness, (3) deepening task comprehension, and (4) multi-component integration. Finally, key challenges and trends in post-training VLA models are identified, establishing a conceptual framework to guide future research. This work delivers both a comprehensive overview of current VLA model post-training methods from a human motor learning perspective and practical insights for VLA model development. (Project website: https://github.com/AoqunJin/Awesome-VLA-Post-Training)
Drones can inspect overhead power lines while they remain energized, significantly simplifying the inspection process. However, localizing a drone relative to all conductors using an onboard LiDAR sensor presents several challenges: (1) conductors provide minimal surface for LiDAR beams limiting the number of conductor points in a scan, (2) not all conductors are consistently detected, and (3) distinguishing LiDAR points corresponding to conductors from other objects, such as trees and pylons, is difficult. This paper proposes an estimation approach that minimizes the error between LiDAR measurements and a single geometric model representing the entire conductor array, rather than tracking individual conductors separately. Experimental results, using data from a power line drone inspection, demonstrate that this method achieves accurate tracking, with a solver converging under 50 ms per frame, even in the presence of partial observations, noise, and outliers. A sensitivity analysis shows that the estimation approach can tolerate up to twice as many outlier points as valid conductors measurements.
Multi-task robotic bimanual manipulation is becoming increasingly popular as it enables sophisticated tasks that require diverse dual-arm collaboration patterns. Compared to unimanual manipulation, bimanual tasks pose challenges to understanding the multi-body spatiotemporal dynamics. An existing method ManiGaussian pioneers encoding the spatiotemporal dynamics into the visual representation via Gaussian world model for single-arm settings, which ignores the interaction of multiple embodiments for dual-arm systems with significant performance drop. In this paper, we propose ManiGaussian++, an extension of ManiGaussian framework that improves multi-task bimanual manipulation by digesting multi-body scene dynamics through a hierarchical Gaussian world model. To be specific, we first generate task-oriented Gaussian Splatting from intermediate visual features, which aims to differentiate acting and stabilizing arms for multi-body spatiotemporal dynamics modeling. We then build a hierarchical Gaussian world model with the leader-follower architecture, where the multi-body spatiotemporal dynamics is mined for intermediate visual representation via future scene prediction. The leader predicts Gaussian Splatting deformation caused by motions of the stabilizing arm, through which the follower generates the physical consequences resulted from the movement of the acting arm. As a result, our method significantly outperforms the current state-of-the-art bimanual manipulation techniques by an improvement of 20.2% in 10 simulated tasks, and achieves 60% success rate on average in 9 challenging real-world tasks. Our code is available at https://github.com/April-Yz/ManiGaussian_Bimanual.
It is difficult to describe in mathematical terms what a neural network trained on data represents. On the other hand, there is a growing mathematical understanding of what neural networks are in principle capable of representing. Feedforward neural networks using the ReLU activation function represent continuous and piecewise linear functions and can approximate many others. The study of their expressivity addresses the question: which ones? Contributing to the available answers, we take the perspective of a neural network as an algorithm. In this analogy, a neural network is programmed constructively, rather than trained from data. An interesting example is a sorting algorithm: we explicitly construct a neural network that sorts its inputs exactly, not approximately, and that, in a sense, has optimal computational complexity if the input dimension is large. Such constructed networks may have several billion parameters. We construct and analyze several other examples, both existing and new. We find that, in these examples, neural networks as algorithms are typically recursive and parallel. Compared to conventional algorithms, ReLU networks are restricted by having to be continuous. Moreover, the depth of recursion is limited by the depth of the network, with deep networks having superior properties over shallow ones.
Standard datasets often present limitations, particularly due to the fixed nature of input data sensors, which makes it difficult to compare methods that actively adjust sensor parameters to suit environmental conditions. This is the case with Automatic-Exposure (AE) methods, which rely on environmental factors to influence the image acquisition process. As a result, AE methods have traditionally been benchmarked in an online manner, rendering experiments non-reproducible. Building on our prior work, we propose a methodology that utilizes an emulator capable of generating images at any exposure time. This approach leverages BorealHDR, a unique multi-exposure stereo dataset, along with its new extension, in which data was acquired along a repeated trajectory at different times of the day to assess the impact of changing illumination. In total, BorealHDR covers 13.4 km over 59 trajectories in challenging lighting conditions. The dataset also includes lidar-inertial-odometry-based maps with pose estimation for each image frame, as well as Global Navigation Satellite System (GNSS) data for comparison. We demonstrate that by using images acquired at various exposure times, we can emulate realistic images with a Root-Mean-Square Error (RMSE) below 1.78% compared to ground truth images. Using this offline approach, we benchmarked eight AE methods, concluding that the classical AE method remains the field's best performer. To further support reproducibility, we provide in-depth details on the development of our backpack acquisition platform, including hardware, electrical components, and performance specifications. Additionally, we share valuable lessons learned from deploying the backpack over more than 25 km across various environments. Our code and dataset are available online at this link: https://github.com/norlab-ulaval/TFR24 BorealHDR
We propose a novel prompt design paradigm that challenges conventional wisdom in large language model (LLM) prompting. While conventional wisdom prioritizes well-crafted instructions and demonstrations for in-context learning (ICL), we show that pruning random demonstrations into seemingly incoherent "gibberish" can remarkably improve performance across diverse tasks. Notably, the "gibberish" always matches or surpasses state-of-the-art automatic prompt optimization techniques, achieving substantial gains regardless of LLM alignment. Nevertheless, discovering an effective pruning strategy is non-trivial, as existing attribution methods and prompt compression algorithms fail to deliver robust results, let alone human intuition. In terms of this, we propose a self-discover prompt optimization framework, PromptQuine, an evolutionary search framework that automatically searches for the pruning strategy by itself using only low-data regimes. Much like the emergent complexity in nature--such as symbiosis and self-organization--arising in response to resource constraints, our framework evolves and refines unconventional yet highly effective prompts by leveraging only the tokens present within the context. We demonstrate its effectiveness across classification, multi-choice question answering, generation and math reasoning tasks across LLMs, while achieving decent runtime efficiency. We hope our findings can guide mechanistic studies on in-context learning, and provide a call to action, to pave the way for more open-ended search algorithms for more effective LLM prompting.
The integration of RGB and thermal data can significantly improve semantic segmentation performance in wild environments for field robots. Nevertheless, multi-source data processing (e.g. Transformer-based approaches) imposes significant computational overhead, presenting challenges for resource-constrained systems. To resolve this critical limitation, we introduced CM-SSM, an efficient RGB-thermal semantic segmentation architecture leveraging a cross-modal state space modeling (SSM) approach. Our framework comprises two key components. First, we introduced a cross-modal 2D-selective-scan (CM-SS2D) module to establish SSM between RGB and thermal modalities, which constructs cross-modal visual sequences and derives hidden state representations of one modality from the other. Second, we developed a cross-modal state space association (CM-SSA) module that effectively integrates global associations from CM-SS2D with local spatial features extracted through convolutional operations. In contrast with Transformer-based approaches, CM-SSM achieves linear computational complexity with respect to image resolution. Experimental results show that CM-SSM achieves state-of-the-art performance on the CART dataset with fewer parameters and lower computational cost. Further experiments on the PST900 dataset demonstrate its generalizability. Codes are available at https://github.com/xiaodonguo/CMSSM.
Understanding the short-term motion of vulnerable road users (VRUs) like pedestrians and cyclists is critical for safe autonomous driving, especially in urban scenarios with ambiguous or high-risk behaviors. While vision-language models (VLMs) have enabled open-vocabulary perception, their utility for fine-grained intent reasoning remains underexplored. Notably, no existing benchmark evaluates multi-class intent prediction in safety-critical situations, To address this gap, we introduce DRAMA-X, a fine-grained benchmark constructed from the DRAMA dataset via an automated annotation pipeline. DRAMA-X contains 5,686 accident-prone frames labeled with object bounding boxes, a nine-class directional intent taxonomy, binary risk scores, expert-generated action suggestions for the ego vehicle, and descriptive motion summaries. These annotations enable a structured evaluation of four interrelated tasks central to autonomous decision-making: object detection, intent prediction, risk assessment, and action suggestion. As a reference baseline, we propose SGG-Intent, a lightweight, training-free framework that mirrors the ego vehicle's reasoning pipeline. It sequentially generates a scene graph from visual input using VLM-backed detectors, infers intent, assesses risk, and recommends an action using a compositional reasoning stage powered by a large language model. We evaluate a range of recent VLMs, comparing performance across all four DRAMA-X tasks. Our experiments demonstrate that scene-graph-based reasoning enhances intent prediction and risk assessment, especially when contextual cues are explicitly modeled.
An ideal traffic simulator replicates the realistic long-term point-to-point trip that a self-driving system experiences during deployment. Prior models and benchmarks focus on closed-loop motion simulation for initial agents in a scene. This is problematic for long-term simulation. Agents enter and exit the scene as the ego vehicle enters new regions. We propose InfGen, a unified next-token prediction model that performs interleaved closed-loop motion simulation and scene generation. InfGen automatically switches between closed-loop motion simulation and scene generation mode. It enables stable long-term rollout simulation. InfGen performs at the state-of-the-art in short-term (9s) traffic simulation, and significantly outperforms all other methods in long-term (30s) simulation. The code and model of InfGen will be released at https://orangesodahub.github.io/InfGen
We introduce a robust framework, RGBTrack, for real-time 6D pose estimation and tracking that operates solely on RGB data, thereby eliminating the need for depth input for such dynamic and precise object pose tracking tasks. Building on the FoundationPose architecture, we devise a novel binary search strategy combined with a render-and-compare mechanism to efficiently infer depth and generate robust pose hypotheses from true-scale CAD models. To maintain stable tracking in dynamic scenarios, including rapid movements and occlusions, RGBTrack integrates state-of-the-art 2D object tracking (XMem) with a Kalman filter and a state machine for proactive object pose recovery. In addition, RGBTrack's scale recovery module dynamically adapts CAD models of unknown scale using an initial depth estimate, enabling seamless integration with modern generative reconstruction techniques. Extensive evaluations on benchmark datasets demonstrate that RGBTrack's novel depth-free approach achieves competitive accuracy and real-time performance, making it a promising practical solution candidate for application areas including robotics, augmented reality, and computer vision. The source code for our implementation will be made publicly available at https://github.com/GreatenAnoymous/RGBTrack.git.
Camera calibration using planar targets has been widely favored, and two types of control points have been mainly considered as measurements: the corners of the checkerboard and the centroid of circles. Since a centroid is derived from numerous pixels, the circular pattern provides more precise measurements than the checkerboard. However, the existing projection model of circle centroids is biased under lens distortion, resulting in low performance. To surmount this limitation, we propose an unbiased projection model of the circular pattern and demonstrate its superior accuracy compared to the checkerboard. Complementing this, we introduce uncertainty into circular patterns to enhance calibration robustness and completeness. Defining centroid uncertainty improves the performance of calibration components, including pattern detection, optimization, and evaluation metrics. We also provide guidelines for performing good camera calibration based on the evaluation metric. The core concept of this approach is to model the boundary points of a two-dimensional shape as a Markov random field, considering its connectivity. The shape distribution is propagated to the centroid uncertainty through an appropriate shape representation based on the Green theorem. Consequently, the resulting framework achieves marked gains in calibration accuracy and robustness. The complete source code and demonstration video are available at https://github.com/chaehyeonsong/discocal.
Recently, robust reinforcement learning (RL) methods designed to handle adversarial input observations have received significant attention, motivated by RL's inherent vulnerabilities. While existing approaches have demonstrated reasonable success, addressing worst-case scenarios over long time horizons requires both minimizing the agent's cumulative rewards for adversaries and training agents to counteract them through alternating learning. However, this process introduces mutual dependencies between the agent and the adversary, making interactions with the environment inefficient and hindering the development of off-policy methods. In this work, we propose a novel off-policy method that eliminates the need for additional environmental interactions by reformulating adversarial learning as a soft-constrained optimization problem. Our approach is theoretically supported by the symmetric property of policy evaluation between the agent and the adversary. The implementation is available at https://github.com/nakanakakosuke/VALT_SAC.
Landslide monitoring is essential for understanding geohazards and mitigating associated risks. However, existing point cloud-based methods typically rely on either geometric or radiometric information and often yield sparse or non-3D displacement estimates. In this paper, we propose a hierarchical partition-based coarse-to-fine approach that fuses 3D point clouds and co-registered RGB images to estimate dense 3D displacement vector fields. We construct patch-level matches using both 3D geometry and 2D image features. These matches are refined via geometric consistency checks, followed by rigid transformation estimation per match. Experimental results on two real-world landslide datasets demonstrate that our method produces 3D displacement estimates with high spatial coverage (79% and 97%) and high accuracy. Deviations in displacement magnitude with respect to external measurements (total station or GNSS observations) are 0.15 m and 0.25 m on the two datasets, respectively, and only 0.07 m and 0.20 m compared to manually derived references. These values are below the average scan resolutions (0.08 m and 0.30 m). Our method outperforms the state-of-the-art method F2S3 in spatial coverage while maintaining comparable accuracy. Our approach offers a practical and adaptable solution for TLS-based landslide monitoring and is extensible to other types of point clouds and monitoring tasks. Our example data and source code are publicly available at https://github.com/zhaoyiww/fusion4landslide.
Neural Image Compression (NIC) has revolutionized image compression with its superior rate-distortion performance and multi-task capabilities, supporting both human visual perception and machine vision tasks. However, its widespread adoption is hindered by substantial computational demands. While existing approaches attempt to address this challenge through module-specific optimizations or pre-defined complexity levels, they lack comprehensive control over computational complexity. We present ABC (Adaptive BayesNet structure learning for computational scalable multi-task image Compression), a novel, comprehensive framework that achieves computational scalability across all NIC components through Bayesian network (BayesNet) structure learning. ABC introduces three key innovations: (i) a heterogeneous bipartite BayesNet (inter-node structure) for managing neural backbone computations; (ii) a homogeneous multipartite BayesNet (intra-node structure) for optimizing autoregressive unit processing; and (iii) an adaptive control module that dynamically adjusts the BayesNet structure based on device capabilities, input data complexity, and downstream task requirements. Experiments demonstrate that ABC enables full computational scalability with better complexity adaptivity and broader complexity control span, while maintaining competitive compression performance. Furthermore, the framework's versatility allows integration with various NIC architectures that employ BayesNet representations, making it a robust solution for ensuring computational scalability in NIC applications. Code is available in https://github.com/worldlife123/cbench_BaSIC.
Detailed captions that accurately reflect the characteristics of a music piece can enrich music databases and drive forward research in music AI. This paper introduces a multi-task music captioning model, SonicVerse, that integrates caption generation with auxiliary music feature detection tasks such as key detection, vocals detection, and more, so as to directly capture both low-level acoustic details as well as high-level musical attributes. The key contribution is a projection-based architecture that transforms audio input into language tokens, while simultaneously detecting music features through dedicated auxiliary heads. The outputs of these heads are also projected into language tokens, to enhance the captioning input. This framework not only produces rich, descriptive captions for short music fragments but also directly enables the generation of detailed time-informed descriptions for longer music pieces, by chaining the outputs using a large-language model. To train the model, we extended the MusicBench dataset by annotating it with music features using MIRFLEX, a modular music feature extractor, resulting in paired audio, captions and music feature data. Experimental results show that incorporating features in this way improves the quality and detail of the generated captions.