Scene graphs have emerged as a structured and serializable environment representation for grounded spatial reasoning with Large Language Models (LLMs). In this work, we propose SG2, an iterative Schema-Guided Scene-Graph reasoning framework based on multi-agent LLMs. The agents are grouped into two modules: a (1) Reasoner module for abstract task planning and graph information queries generation, and a (2) Retriever module for extracting corresponding graph information based on code-writing following the queries. Two modules collaborate iteratively, enabling sequential reasoning and adaptive attention to graph information. The scene graph schema, prompted to both modules, serves to not only streamline both reasoning and retrieval process, but also guide the cooperation between two modules. This eliminates the need to prompt LLMs with full graph data, reducing the chance of hallucination due to irrelevant information. Through experiments in multiple simulation environments, we show that our framework surpasses existing LLM-based approaches and baseline single-agent, tool-based Reason-while-Retrieve strategy in numerical Q&A and planning tasks.

Gaussian Splatting has changed the game for real-time photo-realistic rendering. One of the most popular applications of Gaussian Splatting is to create animatable avatars, known as Gaussian Avatars. Recent works have pushed the boundaries of quality and rendering efficiency but suffer from two main limitations. Either they require expensive multi-camera rigs to produce avatars with free-viewpoint rendering, or they can be trained with a single camera but only rendered at high quality from this fixed viewpoint. An ideal model would be trained using a short monocular video or image from available hardware, such as a webcam, and rendered from any view. To this end, we propose GASP: Gaussian Avatars with Synthetic Priors. To overcome the limitations of existing datasets, we exploit the pixel-perfect nature of synthetic data to train a Gaussian Avatar prior. By fitting this prior model to a single photo or video and fine-tuning it, we get a high-quality Gaussian Avatar, which supports 360° rendering. Our prior is only required for fitting, not inference, enabling real-time applications. Through our method, we obtain high-quality, animatable Avatars from limited data which can be animated and rendered at 70fps on commercial hardware.

Deep neural networks are susceptible to generating overconfident yet erroneous predictions when presented with data beyond known concepts. This challenge underscores the importance of detecting out-of-distribution (OOD) samples in the open world. In this work, we propose a novel feature-space OOD detection score based on class-specific and class-agnostic information. Specifically, the approach utilizes Whitened Linear Discriminant Analysis to project features into two subspaces the discriminative and residual subspaces - for which the in-distribution (ID) classes are maximally separated and closely clustered, respectively. The OOD score is then determined by combining the deviation from the input data to the ID pattern in both subspaces. The efficacy of our method, named WDiscOOD, is verified on the large-scale ImageNet-1k benchmark, with six OOD datasets that cover a variety of distribution shifts. WDiscOOD demonstrates superior performance on deep classifiers with diverse backbone architectures, including CNN and vision transformer. Furthermore, we also show that WDiscOOD more effectively detects novel concepts in representation spaces trained with contrastive objectives, including supervised contrastive loss and multi-modality contrastive loss.

The success of 6-DoF grasp learning with point cloud input is tempered by the computational costs resulting from their unordered nature and pre-processing needs for reducing the point cloud to a manageable size. These properties lead to failure on small objects with low point cloud cardinality. Instead of point clouds, this manuscript explores grasp generation directly from the RGB-D image input. The approach, called Keypoint-GraspNet (KGN), operates in perception space by detecting projected gripper keypoints in the image, then recovering their SE(3) poses with a$\mathrm{P}n\mathrm{P}$algorithm. Training of the network involves a synthetic dataset derived from primitive shape objects with known continuous grasp families. Trained with only single-object synthetic data, Keypoint-GraspNet achieves superior result on our single-object dataset, comparable performance with state-of-art baselines on a multi-object test set, and outperforms the most competitive baseline on small objects. Keypoint-GraspNet is more than 3x faster than tested point cloud methods. Robot experiments show high success rate, demonstrating KGN's practical potential.

We propose an improved keypoint approach for 6-DoF grasp pose synthesis from RGB-D input. Keypoint-based grasp detection from image input demonstrated promising results in a previous study, where the visual information provided by color imagery compensates for noisy or imprecise depth measurements. However, it relies heavily on accurate keypoint prediction in image space. We devise a new grasp generation network that reduces the dependency on precise keypoint estimation. Given an RGB-D input, the network estimates both the grasp pose and the camera-grasp length scale. Re-design of the keypoint output space mitigates the impact of keypoint prediction noise on Perspective-n-Point (PnP) algorithm solutions. Experiments show that the proposed method outperforms the baseline by a large margin, validating its design. Though trained only on simple synthetic objects, our method demonstrates sim-to-real capacity through competitive results in real-world robot experiments.

We consider the task of grasping a target object based on a natural language command query. Previous work primarily focused on localizing the object given the query, which requires a separate grasp detection module to grasp it. The cascaded application of two pipelines incurs errors in overlapping multi-object cases due to ambiguity in the individal outputs. This work proposes a model named Command Grasping Network (CGNet) to directly output command satisficing grasps from RGB image and textual command inputs. A dataset with ground truth (image, command, grasps) tuple is generated based on the VMRD dataset to train the proposed network. Experimental results on the generated test set show that CGNet outperforms a cascaded object-retrieval and grasp detection baseline by a large margin. Three physical experiments demonstrate the functionality and performance of CGNet.

This paper presents a semi-supervised framework for multi-level description learning aiming for robust and accurate camera relocalization across large perception variations. Our proposed network, namely DLSSNet, simultaneously learns weakly-supervised semantic segmentation and local feature description in the hierarchy. Therefore, the augmented descriptors, trained in an end-to-end manner, provide a more stable high-level representation for local feature dis-ambiguity. To facilitate end-to-end semantic description learning, the descriptor segmentation module is proposed to jointly learn semantic descriptors and cluster centers using standard semantic segmentation loss. We show that our model can be easily fine-tuned for domain-specific usage without any further semantic annotations, instead, requiring only 2D-2D pixel correspondences. The learned descriptors, trained with our proposed pipeline, can boost the cross-season localization performance against other state-of-the-arts.