Abstract

This work introduces Talk2BEV, a large vision-language model (LVLM) interface for bird’s-eye view (BEV) maps in autonomous driving contexts. While existing perception systems for autonomous driving scenarios have largely focused on a pre-defined (closed) set of object categories and driving scenarios, Talk2BEV blends recent advances in general-purpose language and vision models with BEV-structured map representations, eliminating the need for task-specific models. This enables a single system to cater to a variety of autonomous driving tasks encompassing visual and spatial reasoning, predicting the intents of traffic actors, and decision-making based on visual cues. We extensively evaluate Talk2BEV on a large number of scene understanding tasks that rely on both the ability to interpret free-form natural language queries and in grounding these queries to the visual context embedded into the language-enhanced BEV map. To enable further research in LVLMs for autonomous driving scenarios, we develop and release Talk2BEV-Bench, a benchmark encompassing 1000 human-annotated BEV scenarios, with more than 20,000 questions and ground-truth responses from the NuScenes dataset.

Abstract

No Results Found

Abstract

With the rise in consumer depth cameras, a wealth of unlabeled RGB-D data has become available. This prompts the question of how to utilize this data for geometric reasoning of scenes. While many RGB-D registration methods rely on geometric and feature-based similarity, we take a different approach. We use cycle-consistent keypoints as salient points to enforce spatial coherence constraints during matching, improving correspondence accuracy. Additionally, we introduce a novel pose block that combines a GRU recurrent unit with transformation synchronization, blending historical and multi-view data. Our approach surpasses previous selfsupervised registration methods on ScanNet and 3DMatch, even outperforming some older supervised methods. We also integrate our components into existing methods, showing their effectiveness.

Abstract

Assistive agents performing household tasks such as making the bed or cooking breakfast often compute and execute actions that accomplish one task at a time. However, efficiency can be improved by anticipating upcoming tasks and computing an action sequence that jointly achieves these tasks. State-of-the-art methods for task anticipation use data-driven deep networks and Large Language Models (LLMs), but they do so at the level of high-level tasks and/or require many training examples. Our framework leverages the generic knowledge of LLMs through a small number of prompts to perform high-level task anticipation, using the anticipated tasks as goals in a classical planning system to compute a sequence of finer-granularity actions that jointly achieve these goals. We ground and evaluate our framework's abilities in realistic scenarios in the VirtualHome environment and demonstrate a 31% reduction in execution time compared with a system that does not consider upcoming tasks.

Abstract

Classical motion planning for robotic manipulation includes a set of general algorithms that aim to minimize a scene-specific cost of executing a given plan. This approach offers remarkable adaptability, as they can be directly used off-the-shelf for any new scene without needing specific training datasets. However, without a prior understanding of what diverse valid trajectories are and without specially designed cost functions for a given scene, the overall solutions tend to have low success rates. While deep-learning-based algorithms tremendously improve success rates, they are much harder to adopt without specialized training datasets. We propose EDMP, an Ensemble-of-costs-guided Diffusion for Motion Planning that aims to combine the strengths of classical and deep-learning-based motion planning. Our diffusion-based network is trained on a set of diverse kinematically valid trajectories. Like classical planning, for any new scene at the time of inference, we compute scene-specific costs such as "collision cost" and guide the diffusion to generate valid trajectories that satisfy the scene-specific constraints. Further, instead of a single cost function that may be insufficient in capturing diversity across scenes, we use an ensemble of costs to guide the diffusion process, significantly improving the success rate compared to classical planners. EDMP performs comparably with SOTA deep-learning-based methods while retaining the generalization capabilities primarily associated with classical planners.

Abstract

Point cloud prediction is an important yet challenging task in the field of autonomous driving. The goal is to predict future point cloud sequences that maintain object structures while accurately representing their temporal motion. These predicted point clouds help in other subsequent tasks like object trajectory estimation for collision avoidance or estimating locations with the least odometry drift. In this work, we present ATPPNet, a novel architecture that predicts future point cloud sequences given a sequence of previous time step point clouds obtained with LiDAR sensor. ATPPNet leverages Conv-LSTM along with channel-wise and spatial attention dually complemented by a 3D-CNN branch for extracting an enhanced spatio-temporal context to recover high quality fidel predictions of future point clouds. We conduct extensive experiments on publicly available datasets and report impressive performance outperforming the existing methods. We also conduct a thorough ablative study of the proposed architecture and provide an application study that highlights the potential of our model for tasks like odometry estimation.

Abstract

Global visual localization in LiDAR-maps, crucial for autonomous driving applications, remains largely unexplored due to the challenging issue of bridging the cross-modal heterogeneity gap. Popular multi-modal learning approach Contrastive Language-Image Pre-Training (CLIP) has popularized contrastive symmetric loss using batch construction technique by applying it to multi-modal domains of text and image. We apply this approach to the domains of 2D image and 3D LiDAR points on the task of cross-modal localization. Our method is explained as follows: A batch of N (image, LiDAR) pairs is constructed so as to predict what is the right match between N X N possible pairings across the batch by jointly training an image encoder and LiDAR encoder to learn a multi-modal embedding space. In this way, the cosine similarity between N positive pairings is maximized, whereas that between the remaining negative pairings is minimized. Finally, over the obtained similarity scores, a symmetric cross-entropy loss is optimized. To the best of our knowledge, this is the first work to apply batched loss approach to a cross-modal setting of image & LiDAR data and also to show Zero-shot transfer in a visual localization setting. We conduct extensive analyses on standard autonomous driving datasets such as KITTI and KITTI-360 datasets. Our method outperforms state-of-the-art recall@1 accuracy on the KITTI-360 dataset by 22.4%, using only perspective images, in contrast to the state-of-the-art approach, which utilizes the more informative fisheye images. Additionally, this superior performance is achieved without resorting to complex architectures. Moreover, we demonstrate the zero-shot capabilities of our model and we beat SOTA by 8% without even training on it. Furthermore, we establish the first benchmark for cross-modal localization on the KITTI dataset.

Abstract

We focus on the problem of LiDAR point cloud based loop detection (or Finding) and closure (LDC) for mobile robots. State-of-the-art (SOTA) methods directly generate learned embeddings from a given point cloud, require large data augmentation, and are not robust to wide viewpoint variations in 6 Degrees-of-Freedom (DOF). Moreover, the absence of strong priors in an unstructured point cloud leads to highly inaccurate LDC. In this original approach, we propose independent roll and pitch canonicalization of point clouds using a common dominant ground plane. We discretize the canonicalized point clouds along the axis perpendicular to the ground plane leads to images similar to digital elevation maps (DEMs), which expose strong spatial priors in the scene. Our experiments show that LDC based on learnt embeddings from such DEMs is not only data efficient but also significantly more robust, and generalizable than the current SOTA. We report an (average precision for loop detection, mean absolute translation/rotation error) improvement of (8.4, 16.7/5.43)% on the KITTI08 sequence, and (11.0, 34.0/25.4)% on GPR10 sequence, over the current SOTA. To further test the robustness of our technique on point clouds in 6-DOF motion we create and opensource a custom dataset called LidarUrbanFly Dataset (LUF) which consists of point clouds obtained from a LiDAR mounted on a quadrotor. More details on our website https://gsc2001.github.io/FinderNet/

Abstract

The Multi-Object Navigation (MultiON) task requires a robot to localize an instance (each) of multiple object classes. It is a fundamental task for an assistive robot in a home or a factory. Existing methods for MultiON have viewed this as a direct extension of Object Navigation (ON), the task of localising an instance of one object class, and are pre-sequenced, i.e., the sequence in which the object classes are to be explored is provided in advance. This is a strong limitation in practical applications characterized by dynamic changes. This paper describes a deep reinforcement learning framework for sequence-agnostic MultiON based on an actor-critic architecture and a suitable reward specification. Our framework leverages past experiences and seeks to reward progress toward individual as well as multiple target object classes. We use photo-realistic scenes from the Gibson benchmark dataset in the AI Habitat 3D simulation environment to experimentally show that our method performs better than a pre-sequenced approach and a state of the art ON method extended to MultiON.

Abstract

Abstract—Building 3D maps of the environment is central to robot navigation, planning, and interaction with objects in a scene. Most existing approaches that integrate semantic concepts with 3D maps largely remain confined to the closed-set setting: they can only reason about a finite set of concepts, pre-defined at training time. Further, these maps can only be queried using class labels, or in more recent work, using text prompts. We address both these issues with ConceptFusion, a scene representation that is: (i) fundamentally open-set, enabling reasoning beyond a closed set of concepts (ii) inherently multimodal, enabling a diverse range of possible queries to the 3D map, from language, to images, to audio, to 3D geometry, all working in concert. ConceptFusion leverages the open-set capabilities of today’s foundation models that have been pretrained on internet-scale data to reason about concepts across modalities such as natural language, images, and audio. We demonstrate that pixel-aligned open-set features can be fused into 3D maps via traditional SLAM and multi-view fusion approaches. This enables effective zero-shot spatial reasoning, not needing any additional training or finetuning, and retains long-tailed concepts better than supervised approaches, outperforming them by more than 40% margin on 3D IoU. We extensively evaluate ConceptFusion on a number of real-world datasets, simulated home environments, a real-world tabletop manipulation task, and an autonomous driving platform. We showcase new avenues for blending foundation models with 3D open-set multimodal mapping. We encourage the reader to view the demos on our project page: https://concept-fusion.github.io