Abhinav Valada

Imen Mahdi, Matteo Cassinelli, Fabien Despinoy et al.

Open-world interactive object search in household environments requires understanding semantic relationships between objects and their surrounding context to guide exploration efficiently. Prior methods either rely on vision-language embeddings similarity, which does not reliably capture task-relevant relational semantics, or large language models (LLMs), which are too slow and costly for real-time deployment. We introduce SCOUT: Scene Graph-Based Exploration with Learned Utility for Open-World Interactive Object Search, a novel method that searches directly over 3D scene graphs by assigning utility scores to rooms, frontiers, and objects using relational exploration heuristics such as room-object containment and object-object co-occurrence. To make this practical without sacrificing open-vocabulary generalization, we propose an offline procedural distillation framework that extracts structured relational knowledge from LLMs into lightweight models for on-robot inference. Furthermore, we present SymSearch, a scalable symbolic benchmark for evaluating semantic reasoning in interactive object search tasks. Extensive evaluations across symbolic and simulation environments show that SCOUT outperforms embedding similarity-based methods and matches LLM-level performance while remaining computationally efficient. Finally, real-world experiments demonstrate effective transfer to physical environments, enabling open-world interactive object search under realistic sensing and navigation constraints.

Nick Heppert, Muhammad Zubair Irshad, Sergey Zakharov et al. · gatech

We present CARTO, a novel approach for reconstructing multiple articulated objects from a single stereo RGB observation. We use implicit object-centric representations and learn a single geometry and articulation decoder for multiple object categories. Despite training on multiple categories, our decoder achieves a comparable reconstruction accuracy to methods that train bespoke decoders separately for each category. Combined with our stereo image encoder we infer the 3D shape, 6D pose, size, joint type, and the joint state of multiple unknown objects in a single forward pass. Our method achieves a 20.4% absolute improvement in mAP 3D IOU50 for novel instances when compared to a two-stage pipeline. Inference time is fast and can run on a NVIDIA TITAN XP GPU at 1 HZ for eight or less objects present. While only trained on simulated data, CARTO transfers to real-world object instances. Code and evaluation data is available at: http://carto.cs.uni-freiburg.de

Nikhil Gosala, Kürsat Petek, Paulo L. J. Drews-Jr et al.

Bird's-Eye-View (BEV) semantic maps have become an essential component of automated driving pipelines due to the rich representation they provide for decision-making tasks. However, existing approaches for generating these maps still follow a fully supervised training paradigm and hence rely on large amounts of annotated BEV data. In this work, we address this limitation by proposing the first self-supervised approach for generating a BEV semantic map using a single monocular image from the frontal view (FV). During training, we overcome the need for BEV ground truth annotations by leveraging the more easily available FV semantic annotations of video sequences. Thus, we propose the SkyEye architecture that learns based on two modes of self-supervision, namely, implicit supervision and explicit supervision. Implicit supervision trains the model by enforcing spatial consistency of the scene over time based on FV semantic sequences, while explicit supervision exploits BEV pseudolabels generated from FV semantic annotations and self-supervised depth estimates. Extensive evaluations on the KITTI-360 dataset demonstrate that our self-supervised approach performs on par with the state-of-the-art fully supervised methods and achieves competitive results using only 1% of direct supervision in the BEV compared to fully supervised approaches. Finally, we publicly release both our code and the BEV datasets generated from the KITTI-360 and Waymo datasets.

Martin Büchner, Jannik Zürn, Ion-George Todoran et al.

Lane graph estimation is an essential and highly challenging task in automated driving and HD map learning. Existing methods using either onboard or aerial imagery struggle with complex lane topologies, out-of-distribution scenarios, or significant occlusions in the image space. Moreover, merging overlapping lane graphs to obtain consistent large-scale graphs remains difficult. To overcome these challenges, we propose a novel bottom-up approach to lane graph estimation from aerial imagery that aggregates multiple overlapping graphs into a single consistent graph. Due to its modular design, our method allows us to address two complementary tasks: predicting ego-respective successor lane graphs from arbitrary vehicle positions using a graph neural network and aggregating these predictions into a consistent global lane graph. Extensive experiments on a large-scale lane graph dataset demonstrate that our approach yields highly accurate lane graphs, even in regions with severe occlusions. The presented approach to graph aggregation proves to eliminate inconsistent predictions while increasing the overall graph quality. We make our large-scale urban lane graph dataset and code publicly available at http://urbanlanegraph.cs.uni-freiburg.de.

Niclas Vödisch, Kürsat Petek, Wolfram Burgard et al.

Operating a robot in the open world requires a high level of robustness with respect to previously unseen environments. Optimally, the robot is able to adapt by itself to new conditions without human supervision, e.g., automatically adjusting its perception system to changing lighting conditions. In this work, we address the task of continual learning for deep learning-based monocular depth estimation and panoptic segmentation in new environments in an online manner. We introduce CoDEPS to perform continual learning involving multiple real-world domains while mitigating catastrophic forgetting by leveraging experience replay. In particular, we propose a novel domain-mixing strategy to generate pseudo-labels to adapt panoptic segmentation. Furthermore, we explicitly address the limited storage capacity of robotic systems by leveraging sampling strategies for constructing a fixed-size replay buffer based on rare semantic class sampling and image diversity. We perform extensive evaluations of CoDEPS on various real-world datasets demonstrating that it successfully adapts to unseen environments without sacrificing performance on previous domains while achieving state-of-the-art results. The code of our work is publicly available at http://codeps.cs.uni-freiburg.de.

Markus Käppeler, Kürsat Petek, Niclas Vödisch et al.

Current state-of-the-art methods for panoptic segmentation require an immense amount of annotated training data that is both arduous and expensive to obtain posing a significant challenge for their widespread adoption. Concurrently, recent breakthroughs in visual representation learning have sparked a paradigm shift leading to the advent of large foundation models that can be trained with completely unlabeled images. In this work, we propose to leverage such task-agnostic image features to enable few-shot panoptic segmentation by presenting Segmenting Panoptic Information with Nearly 0 labels (SPINO). In detail, our method combines a DINOv2 backbone with lightweight network heads for semantic segmentation and boundary estimation. We show that our approach, albeit being trained with only ten annotated images, predicts high-quality pseudo-labels that can be used with any existing panoptic segmentation method. Notably, we demonstrate that SPINO achieves competitive results compared to fully supervised baselines while using less than 0.3% of the ground truth labels, paving the way for learning complex visual recognition tasks leveraging foundation models. To illustrate its general applicability, we further deploy SPINO on real-world robotic vision systems for both outdoor and indoor environments. To foster future research, we make the code and trained models publicly available at http://spino.cs.uni-freiburg.de.

Maximilian Luz, Rohit Mohan, Thomas Nürnberg et al.

Capturing 4D spatiotemporal surroundings is crucial for the safe and reliable operation of robots in dynamic environments. However, most existing methods address only one side of the problem: they either provide coarse geometric tracking via bounding boxes, or detailed 3D structures like voxel-based occupancy that lack explicit temporal association. In this work, we present Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking (LaGS) that advances spatiotemporal scene understanding in a holistic direction. Our approach incorporates camera-based end-to-end tracking with mask-based multi-view panoptic occupancy prediction, and addresses the key challenge of efficiently aggregating multi-view information into 3D voxel grids via a novel latent Gaussian splatting approach. Specifically, we first fuse observations into 3D Gaussians that serve as a sparse point-centric latent representation of the 3D scene, and then splat the aggregated features onto a 3D voxel grid that is decoded by a mask-based segmentation head. We evaluate LaGS on the Occ3D nuScenes and Waymo datasets, achieving state-of-the-art performance for 4D panoptic occupancy tracking. We make our code available at https://lags.cs.uni-freiburg.de/.

Jiarong Wei, Yanxing Chen, Sinuo Song et al.

Learning a single policy that reaches a goal with high geometric precision while interacting safely with nearby agents poses conflicting objectives. Precision favors commitment to a fixed geometric plan, whereas interaction requires immediate deviation when another agent intrudes, causing policies optimized for one objective to often fail at the other. We study this problem in the context of reactive autonomous parking, where multiple vehicles must reach assigned slots with sub-meter terminal accuracy while remaining responsive to neighboring vehicles throughout the maneuver. We propose CoPark, a multi-agent self-play RL approach built on a residual-policy architecture. A precomputed offline plan provides a fixed action prior, while a residual head learns the reactive corrections. The residual policy learns behaviors under self-play, where data and scripting fall short, while the fixed prior holds the slot-frame geometry that pure policies struggle to reach reliably. The key design is a partner-threat-modulated, channel-asymmetric release of the prior. A continuous threat signal shifts authority of the longitudinal channel to the residual head to enable yielding, while the lateral channel remains anchored to the precomputed reference to preserve sub-meter slot alignment. A closed-loop refinement layer corrects residual terminal error from action-grid discretization. We train our policy on six parking lots and evaluate zero-shot on our new reactive-parking benchmark spanning Dragon Lake Parking (DLP) and DeepScenario Open 3D (DSC3D). CoPark achieves ~70-85% success with only 3-6% collision rate, substantially outperforming classical, imitation-learning, and large-scale RL baselines. Importantly, the results demonstrate emergent interaction behaviors such as reverse-yielding, mid-maneuver yielding, tight-corridor passing, and queuing.

Monish R. Nallapareddy, Kshitij Sirohi, Paulo L. J. Drews-Jr et al.

Uncertainty estimation is crucial in safety-critical settings such as automated driving as it provides valuable information for several downstream tasks including high-level decision making and path planning. In this work, we propose EvCenterNet, a novel uncertainty-aware 2D object detection framework using evidential learning to directly estimate both classification and regression uncertainties. To employ evidential learning for object detection, we devise a combination of evidential and focal loss functions for the sparse heatmap inputs. We introduce class-balanced weighting for regression and heatmap prediction to tackle the class imbalance encountered by evidential learning. Moreover, we propose a learning scheme to actively utilize the predicted heatmap uncertainties to improve the detection performance by focusing on the most uncertain points. We train our model on the KITTI dataset and evaluate it on challenging out-of-distribution datasets including BDD100K and nuImages. Our experiments demonstrate that our approach improves the precision and minimizes the execution time loss in relation to the base model.

Fabian Schmidt, Constantin Blessing, Markus Enzweiler et al.

Simultaneous Localization and Mapping (SLAM) is essential for mobile robotics, enabling autonomous navigation in dynamic, unstructured outdoor environments without relying on external positioning systems. These environments pose significant challenges due to variable lighting, weather conditions, and complex terrain. Visual-Inertial SLAM has emerged as a promising solution for robust localization under such conditions. This paper benchmarks several open-source Visual-Inertial SLAM systems, including traditional methods (ORB-SLAM3, VINS-Fusion, OpenVINS, Kimera, and SVO Pro) and learning-based approaches (HFNet-SLAM, AirSLAM), to evaluate their performance in unstructured natural outdoor settings. We focus on the impact of loop closing on localization accuracy and computational demands, providing a comprehensive analysis of these systems' effectiveness in real-world environments and especially their application to embedded systems in outdoor robotics. Our contributions further include an assessment of varying frame rates on localization accuracy and computational load. The findings highlight the importance of loop closing in improving localization accuracy while managing computational resources efficiently, offering valuable insights for optimizing Visual-Inertial SLAM systems for practical outdoor applications in mobile robotics. The dataset and the benchmark code are available under https://github.com/iis-esslingen/vi-slam_lc_benchmark.

D. Adriana Gómez-Rosal, Max Bergau, Georg K. J. Fischer et al.

In today's chemical plants, human field operators perform frequent integrity checks to guarantee high safety standards, and thus are possibly the first to encounter dangerous operating conditions. To alleviate their task, we present a system consisting of an autonomously navigating robot integrated with various sensors and intelligent data processing. It is able to detect methane leaks and estimate its flow rate, detect more general gas anomalies, recognize oil films, localize sound sources and detect failure cases, map the environment in 3D, and navigate autonomously, employing recognition and avoidance of dynamic obstacles. We evaluate our system at a wastewater facility in full working conditions. Our results demonstrate that the system is able to robustly navigate the plant and provide useful information about critical operating conditions.

Fabian Kabus, Julia Hindel, Jelena Bratulić et al. · amazon-science

Recessive dystrophic epidermolysis bullosa (RDEB) is a rare genetic skin disorder for which clinicians greatly benefit from finding similar cases using images and clinical text. However, off-the-shelf foundation models do not reliably capture clinically meaningful features for this heterogeneous, long-tail disease, and structured measurement of agreement with experts is challenging. To address these gaps, we propose evaluating embedding spaces with expert ordinal comparisons (triplet judgments), which are fast to collect and encode implicit clinical similarity knowledge. We further introduce TriDerm, a multimodal framework that learns interpretable wound representations from small cohorts by integrating wound imagery, boundary masks, and expert reports. On the vision side, TriDerm adapts visual foundation models to RDEB using wound-level attention pooling and non-contrastive representation learning. For text, we prompt large language models with comparison queries and recover medically meaningful representations via soft ordinal embeddings (SOE). We show that visual and textual modalities capture complementary aspects of wound phenotype, and that fusing both modalities yields 73.5% agreement with experts, outperforming the best off-the-shelf single-modality foundation model by over 5.6 percentage points. We make the expert annotation tool, model code and representative dataset samples publicly available.

Fabian Schmidt, Karol Fedurko, Markus Enzweiler et al.

While multimodal large language models (MLLMs) provide advanced reasoning for autonomous driving, translating their discrete semantic knowledge into continuous trajectories remains a fundamental challenge. Existing methods often rely on unimodal planning heads that inherently limit their ability to represent multimodal driving behavior. Furthermore, most generative approaches frequently condition on one-hot encoded actions, discarding the nuanced navigational uncertainty critical for complex scenarios. To resolve these limitations, we introduce LAD-Drive, a generative framework that structurally disentangles high-level intention from low-level spatial planning. LAD-Drive employs an action decoder to infer a probabilistic meta-action distribution, establishing an explicit belief state that preserves the nuanced intent typically lost by one-hot encodings. This distribution, fused with the vehicle's kinematic state, conditions an action-aware diffusion decoder that utilizes a truncated denoising process to refine learned motion anchors into safe, kinematically feasible trajectories. Extensive evaluations on the LangAuto benchmark demonstrate that LAD-Drive achieves state-of-the-art results, outperforming competitive baselines by up to 59% in Driving Score while significantly reducing route deviations and collisions. We will publicly release the code and models on https://github.com/iis-esslingen/lad-drive.

Martin Buchner, Abhinav Valada

Online 3D multi-object tracking (MOT) has witnessed significant research interest in recent years, largely driven by demand from the autonomous systems community. However, 3D offline MOT is relatively less explored. Labeling 3D trajectory scene data at a large scale while not relying on high-cost human experts is still an open research question. In this work, we propose Batch3DMOT which follows the tracking-by-detection paradigm and represents real-world scenes as directed, acyclic, and category-disjoint tracking graphs that are attributed using various modalities such as camera, LiDAR, and radar. We present a multi-modal graph neural network that uses a cross-edge attention mechanism mitigating modality intermittence, which translates into sparsity in the graph domain. Additionally, we present attention-weighted convolutions over frame-wise k-NN neighborhoods as suitable means to allow information exchange across disconnected graph components. We evaluate our approach using various sensor modalities and model configurations on the challenging nuScenes and KITTI datasets. Extensive experiments demonstrate that our proposed approach yields an overall improvement of 3.3% in the AMOTA score on nuScenes thereby setting the new state-of-the-art for 3D tracking and further enhancing false positive filtering.

Rohit Mohan, Abhinav Valada

Amodal panoptic segmentation aims to connect the perception of the world to its cognitive understanding. It entails simultaneously predicting the semantic labels of visible scene regions and the entire shape of traffic participant instances, including regions that may be occluded. In this work, we formulate a proposal-free framework that tackles this task as a multi-label and multi-class problem by first assigning the amodal masks to different layers according to their relative occlusion order and then employing amodal instance regression on each layer independently while learning background semantics. We propose the \net architecture that incorporates a shared backbone and an asymmetrical dual-decoder consisting of several modules to facilitate within-scale and cross-scale feature aggregations, bilateral feature propagation between decoders, and integration of global instance-level and local pixel-level occlusion reasoning. Further, we propose the amodal mask refiner that resolves the ambiguity in complex occlusion scenarios by explicitly leveraging the embedding of unoccluded instance masks. Extensive evaluation on the BDD100K-APS and KITTI-360-APS datasets demonstrate that our approach set the new state-of-the-art on both benchmarks.

Fabian Schmidt, Markus Enzweiler, Abhinav Valada

Vision-language models have recently emerged as promising planners for autonomous driving, where success hinges on topology-aware reasoning over spatial structure and dynamic interactions from multimodal input. However, existing models are typically trained without supervision that explicitly encodes these relational dependencies, limiting their ability to infer how agents and other traffic entities influence one another from raw sensor data. In this work, we bridge this gap with a novel model-agnostic method that conditions language-based driving models on structured relational context in the form of traffic scene graphs. We serialize scene graphs at various abstraction levels and formats, and incorporate them into the models via structured prompt templates, enabling a systematic analysis of when and how relational supervision is most beneficial. Extensive evaluations on the public LangAuto benchmark show that scene graph conditioning of state-of-the-art approaches yields large and persistent improvement in driving performance. Notably, we observe up to a 15.6\% increase in driving score for LMDrive and 17.5\% for BEVDriver, indicating that models can better internalize and ground relational priors through scene graph-conditioned training, even without requiring scene graph input at test-time. Code, fine-tuned models, and our scene graph dataset are publicly available at https://github.com/iis-esslingen/GraphPilot.

David Unger, Nikhil Gosala, Varun Ravi Kumar et al.

Most automated driving systems comprise a diverse sensor set, including several cameras, Radars, and LiDARs, ensuring a complete 360°coverage in near and far regions. Unlike Radar and LiDAR, which measure directly in 3D, cameras capture a 2D perspective projection with inherent depth ambiguity. However, it is essential to produce perception outputs in 3D to enable the spatial reasoning of other agents and structures for optimal path planning. The 3D space is typically simplified to the BEV space by omitting the less relevant Z-coordinate, which corresponds to the height dimension.The most basic approach to achieving the desired BEV representation from a camera image is IPM, assuming a flat ground surface. Surround vision systems that are pretty common in new vehicles use the IPM principle to generate a BEV image and to show it on display to the driver. However, this approach is not suited for autonomous driving since there are severe distortions introduced by this too-simplistic transformation method. More recent approaches use deep neural networks to output directly in BEV space. These methods transform camera images into BEV space using geometric constraints implicitly or explicitly in the network. As CNN has more context information and a learnable transformation can be more flexible and adapt to image content, the deep learning-based methods set the new benchmark for BEV transformation and achieve state-of-the-art performance. First, this chapter discusses the contemporary trends of multi-camera-based DNN (deep neural network) models outputting object representations directly in the BEV space. Then, we discuss how this approach can extend to effective sensor fusion and coupling downstream tasks like situation analysis and prediction. Finally, we show challenges and open problems in BEV perception.

Ahmed Rida Sekkat, Rohit Mohan, Oliver Sawade et al.

Unlike humans, who can effortlessly estimate the entirety of objects even when partially occluded, modern computer vision algorithms still find this aspect extremely challenging. Leveraging this amodal perception for autonomous driving remains largely untapped due to the lack of suitable datasets. The curation of these datasets is primarily hindered by significant annotation costs and mitigating annotator subjectivity in accurately labeling occluded regions. To address these limitations, we introduce AmodalSynthDrive, a synthetic multi-task multi-modal amodal perception dataset. The dataset provides multi-view camera images, 3D bounding boxes, LiDAR data, and odometry for 150 driving sequences with over 1M object annotations in diverse traffic, weather, and lighting conditions. AmodalSynthDrive supports multiple amodal scene understanding tasks including the introduced amodal depth estimation for enhanced spatial understanding. We evaluate several baselines for each of these tasks to illustrate the challenges and set up public benchmarking servers. The dataset is available at http://amodalsynthdrive.cs.uni-freiburg.de.

Christopher Lang, Alexander Braun, Lars Schillingmann et al.

Self-supervised feature learning enables perception systems to benefit from the vast raw data recorded by vehicle fleets worldwide. While video-level self-supervised learning approaches have shown strong generalizability on classification tasks, the potential to learn dense representations from sequential data has been relatively unexplored. In this work, we propose TempO, a temporal ordering pretext task for pre-training region-level feature representations for perception tasks. We embed each frame by an unordered set of proposal feature vectors, a representation that is natural for object detection or tracking systems, and formulate the sequential ordering by predicting frame transition probabilities in a transformer-based multi-frame architecture whose complexity scales less than quadratic with respect to the sequence length. Extensive evaluations on the BDD100K, nuImages, and MOT17 datasets show that our TempO pre-training approach outperforms single-frame self-supervised learning methods as well as supervised transfer learning initialization strategies, achieving an improvement of +0.7% in mAP for object detection and +2.0% in the HOTA score for multi-object tracking.

Rohit Mohan, Kiran Kumaraswamy, Juana Valeria Hurtado et al.

Deep learning has led to remarkable strides in scene understanding with panoptic segmentation emerging as a key holistic scene interpretation task. However, the performance of panoptic segmentation is severely impacted in the presence of out-of-distribution (OOD) objects i.e. categories of objects that deviate from the training distribution. To overcome this limitation, we propose Panoptic Out-of Distribution Segmentation for joint pixel-level semantic in-distribution and out-of-distribution classification with instance prediction. We extend two established panoptic segmentation benchmarks, Cityscapes and BDD100K, with out-of-distribution instance segmentation annotations, propose suitable evaluation metrics, and present multiple strong baselines. Importantly, we propose the novel PoDS architecture with a shared backbone, an OOD contextual module for learning global and local OOD object cues, and dual symmetrical decoders with task-specific heads that employ our alignment-mismatch strategy for better OOD generalization. Combined with our data augmentation strategy, this approach facilitates progressive learning of out-of-distribution objects while maintaining in-distribution performance. We perform extensive evaluations that demonstrate that our proposed PoDS network effectively addresses the main challenges and substantially outperforms the baselines. We make the dataset, code, and trained models publicly available at http://pods.cs.uni-freiburg.de.

Julia Hindel, Nikhil Gosala, Kevin Bregler et al.

Perception datasets for agriculture are limited both in quantity and diversity which hinders effective training of supervised learning approaches. Self-supervised learning techniques alleviate this problem, however, existing methods are not optimized for dense prediction tasks in agriculture domains which results in degraded performance. In this work, we address this limitation with our proposed Injected Noise Discriminator (INoD) which exploits principles of feature replacement and dataset discrimination for self-supervised representation learning. INoD interleaves feature maps from two disjoint datasets during their convolutional encoding and predicts the dataset affiliation of the resultant feature map as a pretext task. Our approach enables the network to learn unequivocal representations of objects seen in one dataset while observing them in conjunction with similar features from the disjoint dataset. This allows the network to reason about higher-level semantics of the entailed objects, thus improving its performance on various downstream tasks. Additionally, we introduce the novel Fraunhofer Potato 2022 dataset consisting of over 16,800 images for object detection in potato fields. Extensive evaluations of our proposed INoD pretraining strategy for the tasks of object detection, semantic segmentation, and instance segmentation on the Sugar Beets 2016 and our potato dataset demonstrate that it achieves state-of-the-art performance.

Christopher Lang, Alexander Braun, Lars Schillingmann et al.

Self-supervised multi-object trackers have tremendous potential as they enable learning from raw domain-specific data. However, their re-identification accuracy still falls short compared to their supervised counterparts. We hypothesize that this drawback results from formulating self-supervised objectives that are limited to single frames or frame pairs. Such formulations do not capture sufficient visual appearance variations to facilitate learning consistent re-identification features for autonomous driving when the frame rate is low or object dynamics are high. In this work, we propose a training objective that enables self-supervised learning of re-identification features from multiple sequential frames by enforcing consistent association scores across short and long timescales. We perform extensive evaluations demonstrating that re-identification features trained from longer sequences significantly reduce ID switches on standard autonomous driving datasets compared to existing self-supervised learning methods, which are limited to training on frame pairs. Using our proposed SubCo loss function, we set the new state-of-the-art among self-supervised methods and even perform on par with fully supervised learning methods.

Rohit Mohan, José Arce, Sassan Mokhtar et al.

Safety and efficiency are paramount in healthcare facilities where the lives of patients are at stake. Despite the adoption of robots to assist medical staff in challenging tasks such as complex surgeries, human expertise is still indispensable. The next generation of autonomous healthcare robots hinges on their capacity to perceive and understand their complex and frenetic environments. While deep learning models are increasingly used for this purpose, they require extensive annotated training data which is impractical to obtain in real-world healthcare settings. To bridge this gap, we present Syn-Mediverse, the first hyper-realistic multimodal synthetic dataset of diverse healthcare facilities. Syn-Mediverse contains over \num{48000} images from a simulated industry-standard optical tracking camera and provides more than 1.5M annotations spanning five different scene understanding tasks including depth estimation, object detection, semantic segmentation, instance segmentation, and panoptic segmentation. We demonstrate the complexity of our dataset by evaluating the performance on a broad range of state-of-the-art baselines for each task. To further advance research on scene understanding of healthcare facilities, along with the public dataset we provide an online evaluation benchmark available at \url{http://syn-mediverse.cs.uni-freiburg.de}

Abhijeet Nayak, Daniele Cattaneo, Abhinav Valada

Localization is paramount for autonomous robots. While camera and LiDAR-based approaches have been extensively investigated, they are affected by adverse illumination and weather conditions. Therefore, radar sensors have recently gained attention due to their intrinsic robustness to such conditions. In this paper, we propose RaLF, a novel deep neural network-based approach for localizing radar scans in a LiDAR map of the environment, by jointly learning to address both place recognition and metric localization. RaLF is composed of radar and LiDAR feature encoders, a place recognition head that generates global descriptors, and a metric localization head that predicts the 3-DoF transformation between the radar scan and the map. We tackle the place recognition task by learning a shared embedding space between the two modalities via cross-modal metric learning. Additionally, we perform metric localization by predicting pixel-level flow vectors that align the query radar scan with the LiDAR map. We extensively evaluate our approach on multiple real-world driving datasets and show that RaLF achieves state-of-the-art performance for both place recognition and metric localization. Moreover, we demonstrate that our approach can effectively generalize to different cities and sensor setups than the ones used during training. We make the code and trained models publicly available at http://ralf.cs.uni-freiburg.de.

Christopher Lang, Alexander Braun, Abhinav Valada

Object detection, for the most part, has been formulated in the euclidean space, where euclidean or spherical geodesic distances measure the similarity of an image region to an object class prototype. In this work, we study whether a hyperbolic geometry better matches the underlying structure of the object classification space. We incorporate a hyperbolic classifier in two-stage, keypoint-based, and transformer-based object detection architectures and evaluate them on large-scale, long-tailed, and zero-shot object detection benchmarks. In our extensive experimental evaluations, we observe categorical class hierarchies emerging in the structure of the classification space, resulting in lower classification errors and boosting the overall object detection performance.

Daniel Honerkamp, Tim Welschehold, Abhinav Valada

Despite its importance in both industrial and service robotics, mobile manipulation remains a significant challenge as it requires a seamless integration of end-effector trajectory generation with navigation skills as well as reasoning over long-horizons. Existing methods struggle to control the large configuration space, and to navigate dynamic and unknown environments. In previous work, we proposed to decompose mobile manipulation tasks into a simplified motion generator for the end-effector in task space and a trained reinforcement learning agent for the mobile base to account for kinematic feasibility of the motion. In this work, we introduce Neural Navigation for Mobile Manipulation (N$^2$M$^2$) which extends this decomposition to complex obstacle environments and enables it to tackle a broad range of tasks in real world settings. The resulting approach can perform unseen, long-horizon tasks in unexplored environments while instantly reacting to dynamic obstacles and environmental changes. At the same time, it provides a simple way to define new mobile manipulation tasks. We demonstrate the capabilities of our proposed approach in extensive simulation and real-world experiments on multiple kinematically diverse mobile manipulators. Code and videos are publicly available at http://mobile-rl.cs.uni-freiburg.de.

Maximilian Luz, Rohit Mohan, Ahmed Rida Sekkat et al.

Optical flow estimation is very challenging in situations with transparent or occluded objects. In this work, we address these challenges at the task level by introducing Amodal Optical Flow, which integrates optical flow with amodal perception. Instead of only representing the visible regions, we define amodal optical flow as a multi-layered pixel-level motion field that encompasses both visible and occluded regions of the scene. To facilitate research on this new task, we extend the AmodalSynthDrive dataset to include pixel-level labels for amodal optical flow estimation. We present several strong baselines, along with the Amodal Flow Quality metric to quantify the performance in an interpretable manner. Furthermore, we propose the novel AmodalFlowNet as an initial step toward addressing this task. AmodalFlowNet consists of a transformer-based cost-volume encoder paired with a recurrent transformer decoder which facilitates recurrent hierarchical feature propagation and amodal semantic grounding. We demonstrate the tractability of amodal optical flow in extensive experiments and show its utility for downstream tasks such as panoptic tracking. We make the dataset, code, and trained models publicly available at http://amodal-flow.cs.uni-freiburg.de.

Julia Hindel, Daniele Cattaneo, Abhinav Valada

Semantic segmentation models are typically trained on a fixed set of classes, limiting their applicability in open-world scenarios. Class-incremental semantic segmentation aims to update models with emerging new classes while preventing catastrophic forgetting of previously learned ones. However, existing methods impose strict rigidity on old classes, reducing their effectiveness in learning new incremental classes. In this work, we propose Taxonomy-Oriented Poincaré-regularized Incremental-Class Segmentation (TOPICS) that learns feature embeddings in hyperbolic space following explicit taxonomy-tree structures. This supervision provides plasticity for old classes, updating ancestors based on new classes while integrating new classes at fitting positions. Additionally, we maintain implicit class relational constraints on the geometric basis of the Poincaré ball. This ensures that the latent space can continuously adapt to new constraints while maintaining a robust structure to combat catastrophic forgetting. We also establish eight realistic incremental learning protocols for autonomous driving scenarios, where novel classes can originate from known classes or the background. Extensive evaluations of TOPICS on the Cityscapes and Mapillary Vistas 2.0 benchmarks demonstrate that it achieves state-of-the-art performance. We make the code and trained models publicly available at http://topics.cs.uni-freiburg.de.

Julia Hindel, Simon Bultmann, Houman Masnavi et al.

Self-supervised online traversability estimation enables robots to continuously learn from unlabeled open-world experiences and adapt their navigation behavior toward safe and efficient trajectories. Existing approaches either rely on handcrafted proprioceptive traversability scores, limiting robot-agnosticism, or cluster prior data, preventing online learning. Moreover, many continual learning methods incur substantial memory and computational costs, hindering onboard deployment. We introduce COTRATE, an online learning framework for continuous traversability estimation from multimodal, unlabeled robot experience. Our method first infers robust traversability scores using a robot-agnostic, learning-based online terrain assessment module operating on proprioceptiveand inertial signals. These scores then supervise a visual traversability network through a novel alignment loss that associates visual embeddings with online terrain assessments.To mitigate forgetting during continual learning with minimal overhead, we propose a diversity-aware feature selection strategythat preserves performance using a compact replay memory. We further show that the learned traversability representation supports knowledge transfer across different robot platforms with different locomotion kinematics. We evaluate COTRATE on a dataset of \approx 50,000 images collected with two robotic platforms across 11 outdoor terrains, and benchmark it on navigation tasks in three representative outdoor environments. We make the dataset, code, and trained models publicly available.

Fabian Schmalstieg, Daniel Honerkamp, Tim Welschehold et al.

Existing object-search approaches enable robots to search through free pathways, however, robots operating in unstructured human-centered environments frequently also have to manipulate the environment to their needs. In this work, we introduce a novel interactive multi-object search task in which a robot has to open doors to navigate rooms and search inside cabinets and drawers to find target objects. These new challenges require combining manipulation and navigation skills in unexplored environments. We present HIMOS, a hierarchical reinforcement learning approach that learns to compose exploration, navigation, and manipulation skills. To achieve this, we design an abstract high-level action space around a semantic map memory and leverage the explored environment as instance navigation points. We perform extensive experiments in simulation and the real world that demonstrate that, with accurate perception, the decision making of HIMOS effectively transfers to new environments in a zero-shot manner. It shows robustness to unseen subpolicies, failures in their execution, and different robot kinematics. These capabilities open the door to a wide range of downstream tasks across embodied AI and real-world use cases.

Nicolas Schischka, Nikhil Gosala, B Ravi Kiran et al.

Motion forecasting aims to predict the future trajectories of dynamic agents in the scene, enabling autonomous vehicles to effectively reason about scene evolution. Existing approaches operate under the closed-world regime and assume fixed object taxonomy as well as access to high-quality perception. Therefore, they struggle in real-world settings where perception is imperfect and object taxonomy evolves over time. In this work, we bridge this fundamental gap by introducing open-world motion forecasting, a novel setting in which new object classes are sequentially introduced over time and future object trajectories are estimated directly from camera images. We tackle this setting by proposing the first end-to-end class-incremental motion forecasting framework to mitigate catastrophic forgetting while simultaneously learning to forecast newly introduced classes. When a new class is introduced, our framework employs a pseudo-labeling strategy to first generate motion forecasting pseudo-labels for all known classes which are then processed by a vision-language model to filter inconsistent and over-confident predictions. Parallelly, our approach further mitigates catastrophic forgetting by using a novel replay sampling strategy that leverages query feature variance to sample previous sequences with informative motion patterns. Extensive evaluation on the nuScenes and Argoverse 2 datasets demonstrates that our approach successfully resists catastrophic forgetting and maintains performance on previously learned classes while improving adaptation to novel ones. Further, we demonstrate that our approach supports zero-shot transfer to real-world driving and naturally extends to end-to-end class-incremental planning, enabling continual adaptation of the full autonomous driving system. We provide the code at https://omen.cs.uni-freiburg.de .

Nikhil Gosala, B. Ravi Kiran, Senthil Yogamani et al.

Monocular 3D object tracking aims to estimate temporally consistent 3D object poses across video frames, enabling autonomous agents to reason about scene dynamics. However, existing state-of-the-art approaches are fully supervised and rely on dense 3D annotations over long video sequences, which are expensive to obtain and difficult to scale. In this work, we address this fundamental limitation by proposing the first sparsely supervised framework for monocular 3D object tracking. Our approach decomposes the task into two sequential sub-problems: 2D query matching and 3D geometry estimation. Both components leverage the spatio-temporal consistency of image sequences to augment a sparse set of labeled samples and learn rich 2D and 3D representations of the scene. Leveraging these learned cues, our model automatically generates high-quality 3D pseudolabels across entire videos, effectively transforming sparse supervision into dense 3D track annotations. This enables existing fully-supervised trackers to effectively operate under extreme label sparsity. Extensive experiments on the KITTI and nuScenes datasets demonstrate that our method significantly improves tracking performance, achieving an improvement of up to 15.50 p.p. while using at most four ground truth annotations per track.

Jan Ole von Hartz, Tim Welschehold, Abhinav Valada et al.

Task Parametrized Gaussian Mixture Models (TP-GMM) are a sample-efficient method for learning object-centric robot manipulation tasks. However, there are several open challenges to applying TP-GMMs in the wild. In this work, we tackle three crucial challenges synergistically. First, end-effector velocities are non-Euclidean and thus hard to model using standard GMMs. We thus propose to factorize the robot's end-effector velocity into its direction and magnitude, and model them using Riemannian GMMs. Second, we leverage the factorized velocities to segment and sequence skills from complex demonstration trajectories. Through the segmentation, we further align skill trajectories and hence leverage time as a powerful inductive bias. Third, we present a method to automatically detect relevant task parameters per skill from visual observations. Our approach enables learning complex manipulation tasks from just five demonstrations while using only RGB-D observations. Extensive experimental evaluations on RLBench demonstrate that our approach achieves state-of-the-art performance with 20-fold improved sample efficiency. Our policies generalize across different environments, object instances, and object positions, while the learned skills are reusable.

Juana Valeria Hurtado, Riya Mohan, Abhinav Valada

Forecasting the semantics and 3D structure of scenes is essential for robots to navigate and plan actions safely. Recent methods have explored semantic and panoptic scene forecasting; however, they do not consider the geometry of the scene. In this work, we propose the panoptic-depth forecasting task for jointly predicting the panoptic segmentation and depth maps of unobserved future frames, from monocular camera images. To facilitate this work, we extend the popular KITTI-360 and Cityscapes benchmarks by computing depth maps from LiDAR point clouds and leveraging sequential labeled data. We also introduce a suitable evaluation metric that quantifies both the panoptic quality and depth estimation accuracy of forecasts in a coherent manner. Furthermore, we present two baselines and propose the novel PDcast architecture that learns rich spatio-temporal representations by incorporating a transformer-based encoder, a forecasting module, and task-specific decoders to predict future panoptic-depth outputs. Extensive evaluations demonstrate the effectiveness of PDcast across two datasets and three forecasting tasks, consistently addressing the primary challenges. We make the code publicly available at https://pdcast.cs.uni-freiburg.de.

Max Argus, Abhijeet Nayak, Martin Büchner et al.

Learning-based manipulation policies from image inputs often show weak task transfer capabilities. In contrast, visual servoing methods allow efficient task transfer in high-precision scenarios while requiring only a few demonstrations. In this work, we present a framework that formulates the visual servoing task as graph traversal. Our method not only extends the robustness of visual servoing, but also enables multitask capability based on a few task-specific demonstrations. We construct demonstration graphs by splitting existing demonstrations and recombining them. In order to traverse the demonstration graph in the inference case, we utilize a similarity function that helps select the best demonstration for a specific task. This enables us to compute the shortest path through the graph. Ultimately, we show that recombining demonstrations leads to higher task-respective success. We present extensive simulation and real-world experimental results that demonstrate the efficacy of our approach.

Laura Londoño, Juana Valeria Hurtado, Nora Hertz et al.

Machine learning has significantly enhanced the abilities of robots, enabling them to perform a wide range of tasks in human environments and adapt to our uncertain real world. Recent works in various machine learning domains have highlighted the importance of accounting for fairness to ensure that these algorithms do not reproduce human biases and consequently lead to discriminatory outcomes. With robot learning systems increasingly performing more and more tasks in our everyday lives, it is crucial to understand the influence of such biases to prevent unintended behavior toward certain groups of people. In this work, we present the first survey on fairness in robot learning from an interdisciplinary perspective spanning technical, ethical, and legal challenges. We propose a taxonomy for sources of bias and the resulting types of discrimination due to them. Using examples from different robot learning domains, we examine scenarios of unfair outcomes and strategies to mitigate them. We present early advances in the field by covering different fairness definitions, ethical and legal considerations, and methods for fair robot learning. With this work, we aim to pave the road for groundbreaking developments in fair robot learning.

Sharang Kaul, Simon Bultmann, Mario Berk et al.

Criticality metrics such as time-to-collision (TTC) quantify collision urgency but conflate the consequences of false-positive (FP) and false-negative (FN) perception errors. We propose two novel effort-based metrics: False Speed Reduction (FSR), the cumulative velocity loss from persistent phantom detections, and Maximum Deceleration Rate (MDR), the peak braking demand from missed objects under a constant-acceleration model. These longitudinal metrics are complemented by Lateral Evasion Acceleration (LEA), adapted from prior lateral evasion kinematics and coupled with reachability-based collision timing to quantify the minimum steering effort to avoid a predicted collision. A reachability-based ellipsoidal collision filter ensures only dynamically plausible threats are scored, with frame-level matching and track-level aggregation. Evaluation of different perception pipelines on nuScenes and Argoverse~2 shows that 65-93% of errors are non-critical, and Spearman correlation analysis confirms that all three metrics capture safety-relevant information inaccessible to established time-based, deceleration-based, or normalized criticality measures, enabling targeted mining of the most critical perception failures.

Tim Engelbracht, René Zurbrügg, Matteo Wohlrapp et al.

We present a dataset for force-grounded, cross-view articulated manipulation that couples what is seen with what is done and what is felt during real human interaction. The dataset contains 3048 sequences across 381 articulated objects in 38 environments. Each object is operated in four embodiments - (i) human hand, (ii) human hand with a wrist-mounted camera, (iii) handheld UMI gripper, and (iv) a custom Hoi! gripper, where the tool embodiment provides end-effector forces and tactile sensing. Our dataset offers a holistic view of interaction understanding from video, enabling researchers to evaluate how well methods transfer between human and robotic viewpoints, but also investigate underexplored modalities such as interaction forces. The Project Website can be found at https://timengelbracht.github.io/Hoi-Dataset-Website/.

Daniel Honerkamp, Harsh Mahesheka, Jan Ole von Hartz et al.

Demonstration data plays a key role in learning complex behaviors and training robotic foundation models. While effective control interfaces exist for static manipulators, data collection remains cumbersome and time intensive for mobile manipulators due to their large number of degrees of freedom. While specialized hardware, avatars, or motion tracking can enable whole-body control, these approaches are either expensive, robot-specific, or suffer from the embodiment mismatch between robot and human demonstrator. In this work, we present MoMa-Teleop, a novel teleoperation method that infers end-effector motions from existing interfaces and delegates the base motions to a previously developed reinforcement learning agent, leaving the operator to focus fully on the task-relevant end-effector motions. This enables whole-body teleoperation of mobile manipulators with no additional hardware or setup costs via standard interfaces such as joysticks or hand guidance. Moreover, the operator is not bound to a tracked workspace and can move freely with the robot over spatially extended tasks. We demonstrate that our approach results in a significant reduction in task completion time across a variety of robots and tasks. As the generated data covers diverse whole-body motions without embodiment mismatch, it enables efficient imitation learning. By focusing on task-specific end-effector motions, our approach learns skills that transfer to unseen settings, such as new obstacles or changed object positions, from as little as five demonstrations. We make code and videos available at https://moma-teleop.cs.uni-freiburg.de.

Mohamed Abdelsamad, Michael Ulrich, Bin Yang et al.

Recent advances in self-supervised learning (SSL) have shown tremendous potential for learning 3D point cloud representations without human annotations. However, SSL for 3D point clouds still faces critical challenges due to irregular geometry, shortcut-prone reconstruction, and unbalanced semantics distribution. In this work, we propose DOS (Distilling Observable Softmaps), a novel SSL framework that self-distills semantic relevance softmaps only at observable (unmasked) points. This strategy prevents information leakage from masked regions and provides richer supervision than discrete token-to-prototype assignments. To address the challenge of unbalanced semantics in an unsupervised setting, we introduce Zipfian prototypes and incorporate them using a modified Sinkhorn-Knopp algorithm, Zipf-Sinkhorn, which enforces a power-law prior over prototype usage and modulates the sharpness of the target softmap during training. DOS outperforms current state-of-the-art methods on semantic segmentation and 3D object detection across multiple benchmarks, including nuScenes, Waymo, SemanticKITTI, ScanNet, and ScanNet200, without relying on extra data or annotations. Our results demonstrate that observable-point softmaps distillation offers a scalable and effective paradigm for learning robust 3D representations.

Fabian Kabus, Maren Hackenberg, Julia Hindel et al.

While artificial neural networks excel in unsupervised learning of non-sparse structure, classical statistical regression techniques offer better interpretability, in particular when sparseness is enforced by $\ell_1$ regularization, enabling identification of which factors drive observed dynamics. We investigate how these two types of approaches can be optimally combined, exemplarily considering two-photon calcium imaging data where sparse autoregressive dynamics are to be extracted. We propose embedding a vector autoregressive (VAR) model as an interpretable regression technique into a convolutional autoencoder, which provides dimension reduction for tractable temporal modeling. A skip connection separately addresses non-sparse static spatial information, selectively channeling sparse structure into the $\ell_1$-regularized VAR. $\ell_1$-estimation of regression parameters is enabled by differentiating through the piecewise linear solution path. This is contrasted with approaches where the autoencoder does not adapt to the VAR model. Having an embedded statistical model also enables a testing approach for comparing temporal sequences from the same observational unit. Additionally, contribution maps visualize which spatial regions drive the learned dynamics.

Seyed Mahdi B. Azad, Jasper Hoffmann, Iman Nematollahi et al.

Forward-backward (FB) representations provide a powerful framework for learning the successor representation (SR) in continuous spaces by enforcing a low-rank factorization. However, a fundamental spectral mismatch often exists between the high-rank transition dynamics of continuous environments and the low-rank bottleneck of the FB architecture, making accurate low-rank representation learning difficult. In this work, we analyze temporal abstraction as a mechanism to mitigate this mismatch. By characterizing the spectral properties of the transition operator, we show that temporal abstraction acts as a low-pass filter that suppresses high-frequency spectral components. This suppression reduces the effective rank of the induced SR while preserving a formal bound on the resulting value function error. Empirically, we show that this alignment is a key factor for stable FB learning, particularly at high discount factors where bootstrapping becomes error-prone. Our results identify temporal abstraction as a principled mechanism for shaping the spectral structure of the underlying MDP and enabling effective long-horizon representations in continuous control.

Jan Ole von Hartz, Eugenio Chisari, Tim Welschehold et al.

In recent years, policy learning methods using either reinforcement or imitation have made significant progress. However, both techniques still suffer from being computationally expensive and requiring large amounts of training data. This problem is especially prevalent in real-world robotic manipulation tasks, where access to ground truth scene features is not available and policies are instead learned from raw camera observations. In this paper, we demonstrate the efficacy of learning image keypoints via the Dense Correspondence pretext task for downstream policy learning. Extending prior work to challenging multi-object scenes, we show that our model can be trained to deal with important problems in representation learning, primarily scale-invariance and occlusion. We evaluate our approach on diverse robot manipulation tasks, compare it to other visual representation learning approaches, and demonstrate its flexibility and effectiveness for sample-efficient policy learning.

Aron Distelzweig, Eitan Kosman, Andreas Look et al.

Forecasting the future trajectories of surrounding agents is crucial for autonomous vehicles to ensure safe, efficient, and comfortable route planning. While model ensembling has improved prediction accuracy in various fields, its application in trajectory prediction is limited due to the multi-modal nature of predictions. In this paper, we propose a novel sampling method applicable to trajectory prediction based on the predictions of multiple models. We first show that conventional sampling based on predicted probabilities can degrade performance due to missing alignment between models. To address this problem, we introduce a new method that generates optimal trajectories from a set of neural networks, framing it as a risk minimization problem with a variable loss function. By using state-of-the-art models as base learners, our approach constructs diverse and effective ensembles for optimal trajectory sampling. Extensive experiments on the nuScenes prediction dataset demonstrate that our method surpasses current state-of-the-art techniques, achieving top ranks on the leaderboard. We also provide a comprehensive empirical study on ensembling strategies, offering insights into their effectiveness. Our findings highlight the potential of advanced ensembling techniques in trajectory prediction, significantly improving predictive performance and paving the way for more reliable predicted trajectories.

Simon Bultmann, Daniele Cattaneo, Abhinav Valada

Accurate camera-LiDAR calibration is a prerequisite for robust multi-modal perception in robotics. Recent target-less approaches based on deep point correspondences achieve remarkable performance for extrinsic calibration but assume rectified images with known intrinsics. In this work, we overcome this limitation and present the first fully target-less pipeline that jointly estimates camera intrinsics (pinhole model with radial-tangential distortion) and camera-LiDAR extrinsics with deep pixel-point correspondences. Our approach extends deep correspondence-based calibration by (i) automatic intrinsic initialization via structure-from-motion, (ii) generalizing camera-LiDAR matching to raw images with unknown intrinsics including distortion, and (iii) tightly coupling correspondence estimation with joint nonlinear optimization over both intrinsics and extrinsics. We evaluate our method on the KITTI dataset with unseen camera-LiDAR pairs and demonstrate that joint calibration achieves improved extrinsic accuracy while additionally recovering accurate intrinsics.

Jonas Schramm, Niclas Vödisch, Kürsat Petek et al.

Semantic scene segmentation from a bird's-eye-view (BEV) perspective plays a crucial role in facilitating planning and decision-making for mobile robots. Although recent vision-only methods have demonstrated notable advancements in performance, they often struggle under adverse illumination conditions such as rain or nighttime. While active sensors offer a solution to this challenge, the prohibitively high cost of LiDARs remains a limiting factor. Fusing camera data with automotive radars poses a more inexpensive alternative but has received less attention in prior research. In this work, we aim to advance this promising avenue by introducing BEVCar, a novel approach for joint BEV object and map segmentation. The core novelty of our approach lies in first learning a point-based encoding of raw radar data, which is then leveraged to efficiently initialize the lifting of image features into the BEV space. We perform extensive experiments on the nuScenes dataset and demonstrate that BEVCar outperforms the current state of the art. Moreover, we show that incorporating radar information significantly enhances robustness in challenging environmental conditions and improves segmentation performance for distant objects. To foster future research, we provide the weather split of the nuScenes dataset used in our experiments, along with our code and trained models at http://bevcar.cs.uni-freiburg.de.

Sai Prasanna, Daniel Honerkamp, Kshitij Sirohi et al.

Embodied AI has made significant progress acting in unexplored environments. However, tasks such as object search have largely focused on efficient policy learning. In this work, we identify several gaps in current search methods: They largely focus on dated perception models, neglect temporal aggregation, and transfer from ground truth directly to noisy perception at test time, without accounting for the resulting overconfidence in the perceived state. We address the identified problems through calibrated perception probabilities and uncertainty across aggregation and found decisions, thereby adapting the models for sequential tasks. The resulting methods can be directly integrated with pretrained models across a wide family of existing search approaches at no additional training cost. We perform extensive evaluations of aggregation methods across both different semantic perception models and policies, confirming the importance of calibrated uncertainties in both the aggregation and found decisions. We make the code and trained models available at https://semantic-search.cs.uni-freiburg.de.

Liudi Yang, George Eskandar, Fengyi Shen et al.

We address the challenge of novel view synthesis from only two input images under large viewpoint changes. Existing regression-based methods lack the capacity to reconstruct unseen regions, while camera-guided diffusion models often deviate from intended trajectories due to noisy point cloud projections or insufficient conditioning from camera poses. To address these issues, we propose ConfCtrl, a confidence-aware video interpolation framework that enables diffusion models to follow prescribed camera poses while completing unseen regions. ConfCtrl initializes the diffusion process by combining a confidence-weighted projected point cloud latent with noise as the conditioning input. It then applies a Kalman-inspired predict-update mechanism, treating the projected point cloud as a noisy measurement and using learned residual corrections to balance pose-driven predictions with noisy geometric observations. This allows the model to rely on reliable projections while down-weighting uncertain regions, yielding stable, geometry-aware generation. Experiments on multiple datasets show that ConfCtrl produces geometrically consistent and visually plausible novel views, effectively reconstructing occluded regions under large viewpoint changes.

Martin Büchner, Adrian Röfer, Tim Engelbracht et al.

Semantics has enabled 3D scene understanding and affordance-driven object interaction. However, robots operating in real-world environments face a critical limitation: they cannot anticipate how objects move. Long-horizon mobile manipulation requires closing the gap between semantics, geometry, and kinematics. In this work, we present MoMa-SG, a novel framework for building semantic-kinematic 3D scene graphs of articulated scenes containing a myriad of interactable objects. Given RGB-D sequences containing multiple object articulations, we temporally segment object interactions and infer object motion using occlusion-robust point tracking. We then lift point trajectories into 3D and estimate articulation models using a novel unified twist estimation formulation that robustly estimates revolute and prismatic joint parameters in a single optimization pass. Next, we associate objects with estimated articulations and detect contained objects by reasoning over parent-child relations at identified opening states. We also introduce the novel Arti4D-Semantic dataset, which uniquely combines hierarchical object semantics including parent-child relation labels with object axis annotations across 62 in-the-wild RGB-D sequences containing 600 object interactions and three distinct observation paradigms. We extensively evaluate the performance of MoMa-SG on two datasets and ablate key design choices of our approach. In addition, real-world experiments on both a quadruped and a mobile manipulator demonstrate that our semantic-kinematic scene graphs enable robust manipulation of articulated objects in everyday home environments. We provide code and data at: https://momasg.cs.uni-freiburg.de.

Liudi Yang, Yang Bai, George Eskandar et al.

We present a method to generate video-action pairs that follow text instructions, starting from an initial image observation and the robot's joint states. Our approach automatically provides action labels for video diffusion models, overcoming the common lack of action annotations and enabling their full use for robotic policy learning. Existing methods either adopt two-stage pipelines, which limit tightly coupled cross-modal information sharing, or rely on adapting a single-modal diffusion model for a joint distribution that cannot fully leverage pretrained video knowledge. To overcome these limitations, we (1) extend a pretrained video diffusion model with a parallel, dedicated action diffusion model that preserves pretrained knowledge, (2) introduce a Bridge Attention mechanism to enable effective cross-modal interaction, and (3) design an action refinement module to convert coarse actions into precise controls for low-resolution datasets. Extensive evaluations on multiple public benchmarks and real-world datasets demonstrate that our method generates higher-quality videos, more accurate actions, and significantly outperforms existing baselines, offering a scalable framework for leveraging large-scale video data for robotic learning.