Jiazhao Zhang

Qiyu Dai, Yan Zhu, Yiran Geng et al. · pku

In this work, we tackle 6-DoF grasp detection for transparent and specular objects, which is an important yet challenging problem in vision-based robotic systems, due to the failure of depth cameras in sensing their geometry. We, for the first time, propose a multiview RGB-based 6-DoF grasp detection network, GraspNeRF, that leverages the generalizable neural radiance field (NeRF) to achieve material-agnostic object grasping in clutter. Compared to the existing NeRF-based 3-DoF grasp detection methods that rely on densely captured input images and time-consuming per-scene optimization, our system can perform zero-shot NeRF construction with sparse RGB inputs and reliably detect 6-DoF grasps, both in real-time. The proposed framework jointly learns generalizable NeRF and grasp detection in an end-to-end manner, optimizing the scene representation construction for the grasping. For training data, we generate a large-scale photorealistic domain-randomized synthetic dataset of grasping in cluttered tabletop scenes that enables direct transfer to the real world. Our extensive experiments in synthetic and real-world environments demonstrate that our method significantly outperforms all the baselines in all the experiments while remaining in real-time. Project page can be found at https://pku-epic.github.io/GraspNeRF

Jiangran Lyu, Kai Liu, Xuheng Zhang et al.

Recent robot foundation models largely rely on large-scale behavior cloning, which imitates expert actions but discards transferable dynamics knowledge embedded in heterogeneous embodied data. While the Unified World Model (UWM) formulation has the potential to leverage such diverse data, existing instantiations struggle to scale to foundation-level due to coarse data usage and fragmented datasets. We introduce LDA-1B, a robot foundation model that scales through universal embodied data ingestion by jointly learning dynamics, policy, and visual forecasting, assigning distinct roles to data of varying quality. To support this regime at scale, we assemble and standardize EI-30k, an embodied interaction dataset comprising over 30k hours of human and robot trajectories in a unified format. Scalable dynamics learning over such heterogeneous data is enabled by prediction in a structured DINO latent space, which avoids redundant pixel-space appearance modeling. Complementing this representation, LDA-1B employs a multi-modal diffusion transformer to handle asynchronous vision and action streams, enabling stable training at the 1B-parameter scale. Experiments in simulation and the real world show LDA-1B outperforms prior methods (e.g., $π_{0.5}$) by up to 21\%, 48\%, and 23\% on contact-rich, dexterous, and long-horizon tasks, respectively. Notably, LDA-1B enables data-efficient fine-tuning, gaining 10\% by leveraging 30\% low-quality trajectories typically harmful and discarded.

Jiayi Chen, Mi Yan, Jiazhao Zhang et al.

In this work, we tackle the challenging task of jointly tracking hand object pose and reconstructing their shapes from depth point cloud sequences in the wild, given the initial poses at frame 0. We for the first time propose a point cloud based hand joint tracking network, HandTrackNet, to estimate the inter-frame hand joint motion. Our HandTrackNet proposes a novel hand pose canonicalization module to ease the tracking task, yielding accurate and robust hand joint tracking. Our pipeline then reconstructs the full hand via converting the predicted hand joints into a template-based parametric hand model MANO. For object tracking, we devise a simple yet effective module that estimates the object SDF from the first frame and performs optimization-based tracking. Finally, a joint optimization step is adopted to perform joint hand and object reasoning, which alleviates the occlusion-induced ambiguity and further refines the hand pose. During training, the whole pipeline only sees purely synthetic data, which are synthesized with sufficient variations and by depth simulation for the ease of generalization. The whole pipeline is pertinent to the generalization gaps and thus directly transferable to real in-the-wild data. We evaluate our method on two real hand object interaction datasets, e.g. HO3D and DexYCB, without any finetuning. Our experiments demonstrate that the proposed method significantly outperforms the previous state-of-the-art depth-based hand and object pose estimation and tracking methods, running at a frame rate of 9 FPS.

Jiazhao Zhang, Liu Dai, Fanpeng Meng et al.

Object goal navigation (ObjectNav) in unseen environments is a fundamental task for Embodied AI. Agents in existing works learn ObjectNav policies based on 2D maps, scene graphs, or image sequences. Considering this task happens in 3D space, a 3D-aware agent can advance its ObjectNav capability via learning from fine-grained spatial information. However, leveraging 3D scene representation can be prohibitively unpractical for policy learning in this floor-level task, due to low sample efficiency and expensive computational cost. In this work, we propose a framework for the challenging 3D-aware ObjectNav based on two straightforward sub-policies. The two sub-polices, namely corner-guided exploration policy and category-aware identification policy, simultaneously perform by utilizing online fused 3D points as observation. Through extensive experiments, we show that this framework can dramatically improve the performance in ObjectNav through learning from 3D scene representation. Our framework achieves the best performance among all modular-based methods on the Matterport3D and Gibson datasets, while requiring (up to 30x) less computational cost for training.

Jiazhao Zhang, Nandiraju Gireesh, Jilong Wang et al.

Mobile manipulation constitutes a fundamental task for robotic assistants and garners significant attention within the robotics community. A critical challenge inherent in mobile manipulation is the effective observation of the target while approaching it for grasping. In this work, we propose a graspability-aware mobile manipulation approach powered by an online grasping pose fusion framework that enables a temporally consistent grasping observation. Specifically, the predicted grasping poses are online organized to eliminate the redundant, outlier grasping poses, which can be encoded as a grasping pose observation state for reinforcement learning. Moreover, on-the-fly fusing the grasping poses enables a direct assessment of graspability, encompassing both the quantity and quality of grasping poses.

Yijie Tang, Jiazhao Zhang, Zhinan Yu et al.

We introduce MIPS-Fusion, a robust and scalable online RGB-D reconstruction method based on a novel neural implicit representation -- multi-implicit-submap. Different from existing neural RGB-D reconstruction methods lacking either flexibility with a single neural map or scalability due to extra storage of feature grids, we propose a pure neural representation tackling both difficulties with a divide-and-conquer design. In our method, neural submaps are incrementally allocated alongside the scanning trajectory and efficiently learned with local neural bundle adjustments. The submaps can be refined individually in a back-end optimization and optimized jointly to realize submap-level loop closure. Meanwhile, we propose a hybrid tracking approach combining randomized and gradient-based pose optimizations. For the first time, randomized optimization is made possible in neural tracking with several key designs to the learning process, enabling efficient and robust tracking even under fast camera motions. The extensive evaluation demonstrates that our method attains higher reconstruction quality than the state of the arts for large-scale scenes and under fast camera motions.

Qiuyue Wang, Mingsheng Li, Jian Guan et al.

Embodied intelligence is often studied through specialized models for individual tasks such as manipulation or navigation, resulting in fragmented capabilities and limited generalization across tasks, environments, and robot embodiments. In this work, we study whether heterogeneous embodied decision-making problems can be unified within a single vision-language-action model. We present Qwen-VLA, a unified embodied foundation model that extends Qwen's vision-language modeling stack from perception, understanding, and reasoning to continuous action and trajectory generation through a DiT-based action decoder. Qwen-VLA is trained with a large-scale joint pretraining recipe over diverse data sources, including robotics manipulation trajectories, human egocentric demonstrations, synthetic simulation data, vision-and-language navigation data, trajectory-centric supervision, and auxiliary vision-language data. To support multiple robot platforms, we introduce embodiment-aware prompt conditioning, where robot-specific textual descriptions specify the current embodiment and control convention. We further cast manipulation, navigation, and trajectory prediction into a unified action-and-trajectory prediction framework, enabling transferable visual grounding, spatial reasoning, and continuous action generation across robot morphologies, task families, and environments. Experiments on manipulation, navigation, and trajectory-centric benchmarks show consistent multi-task performance and out-of-distribution generalization under variations in scene layout, background, lighting, object configuration, and robot embodiment. Qwen-VLA-Instruct achieves 97.9% on LIBERO, 73.7% on Simpler-WidowX, 86.1%/87.2% on RoboTwin-Easy/Hard, 69.0% OSR on R2R, 59.6% SR on RxR, 76.9% average OOD success in real-world ALOHA experiments, and 26.6% zero-shot success on DOMINO dynamic manipulation.

Songlin Wei, Jiazhao Zhang, Yang Wang et al.

Neural radiance fields with stochasticity have garnered significant interest by enabling the sampling of plausible radiance fields and quantifying uncertainty for downstream tasks. Existing works rely on the independence assumption of points in the radiance field or the pixels in input views to obtain tractable forms of the probability density function. However, this assumption inadvertently impacts performance when dealing with intricate geometry and texture. In this work, we propose an independence-assumption-free probabilistic neural radiance field based on Flow-GAN. By combining the generative capability of adversarial learning and the powerful expressivity of normalizing flow, our method explicitly models the density-radiance distribution of the whole scene. We represent our probabilistic NeRF as a mean-shifted probabilistic residual neural model. Our model is trained without an explicit likelihood function, thereby avoiding the independence assumption. Specifically, We downsample the training images with different strides and centers to form fixed-size patches which are used to train the generator with patch-based adversarial learning. Through extensive experiments, our method demonstrates state-of-the-art performance by predicting lower rendering errors and more reliable uncertainty on both synthetic and real-world datasets.

Ziyuan Xia, Jingyi Xu, Chong Cui et al.

Training embodied AI agents depends critically on the visual fidelity of simulation environments and the ability to model dynamic humans. Current simulators rely on mesh-based rasterization with limited visual realism, and their support for dynamic human avatars, where available, is constrained to mesh representations, hindering agent generalization to human-populated real-world scenarios. We present Habitat-GS, a navigation-centric embodied AI simulator extended from Habitat-Sim that integrates 3D Gaussian Splatting scene rendering and drivable gaussian avatars while maintaining full compatibility with the Habitat ecosystem. Our system implements a 3DGS renderer for real-time photorealistic rendering and supports scalable 3DGS asset import from diverse sources. For dynamic human modeling, we introduce a gaussian avatar module that enables each avatar to simultaneously serve as a photorealistic visual entity and an effective navigation obstacle, allowing agents to learn human-aware behaviors in realistic settings. Experiments on point-goal navigation demonstrate that agents trained on 3DGS scenes achieve stronger cross-domain generalization, with mixed-domain training being the most effective strategy. Evaluations on avatar-aware navigation further confirm that gaussian avatars enable effective human-aware navigation. Finally, performance benchmarks validate the system's scalability across varying scene complexity and avatar counts.

Jiahang Liu, Tianyu Xu, Jiawei Chen et al.

Recent embodied navigation approaches leveraging Vision-Language Models (VLMs) demonstrate strong generalization in versatile Vision-Language Navigation (VLN). However, reliable path planning in complex environments remains challenging due to insufficient spatial awareness. In this work, we introduce SPAN-Nav, an end-to-end foundation model designed to infuse embodied navigation with universal 3D spatial awareness using RGB video streams. SPAN-Nav extracts spatial priors across diverse scenes through an occupancy prediction task on extensive indoor and outdoor environments. To mitigate the computational burden, we introduce a compact representation for spatial priors, finding that a single token is sufficient to encapsulate the coarse-grained cues essential for navigation tasks. Furthermore, inspired by the Chain-of-Thought (CoT) mechanism, SPAN-Nav utilizes this single spatial token to explicitly inject spatial cues into action reasoning through an end-to end framework. Leveraging multi-task co-training, SPAN-Nav captures task-adaptive cues from generalized spatial priors, enabling robust spatial awareness to generalize even to the task lacking explicit spatial supervision. To support comprehensive spatial learning, we present a massive dataset of 4.2 million occupancy annotations that covers both indoor and outdoor scenes across multi-type navigation tasks. SPAN-Nav achieves state-of-the-art performance across three benchmarks spanning diverse scenarios and varied navigation tasks. Finally, real-world experiments validate the robust generalization and practical reliability of our approach across complex physical scenarios.

Jiazhao Zhang, Kunyu Wang, Rongtao Xu et al.

Vision-and-language navigation (VLN) stands as a key research problem of Embodied AI, aiming at enabling agents to navigate in unseen environments following linguistic instructions. In this field, generalization is a long-standing challenge, either to out-of-distribution scenes or from Sim to Real. In this paper, we propose NaVid, a video-based large vision language model (VLM), to mitigate such a generalization gap. NaVid makes the first endeavor to showcase the capability of VLMs to achieve state-of-the-art level navigation performance without any maps, odometers, or depth inputs. Following human instruction, NaVid only requires an on-the-fly video stream from a monocular RGB camera equipped on the robot to output the next-step action. Our formulation mimics how humans navigate and naturally gets rid of the problems introduced by odometer noises, and the Sim2Real gaps from map or depth inputs. Moreover, our video-based approach can effectively encode the historical observations of robots as spatio-temporal contexts for decision making and instruction following. We train NaVid with 510k navigation samples collected from continuous environments, including action-planning and instruction-reasoning samples, along with 763k large-scale web data. Extensive experiments show that NaVid achieves state-of-the-art performance in simulation environments and the real world, demonstrating superior cross-dataset and Sim2Real transfer. We thus believe our proposed VLM approach plans the next step for not only the navigation agents but also this research field.

Jiahang Liu, Yuanxing Duan, Jiazhao Zhang et al.

Simulating realistic environments for robots is widely recognized as a critical challenge in robot learning, particularly in terms of rendering and physical simulation. This challenge becomes even more pronounced in navigation tasks, where trajectories often extend across multiple rooms or entire floors. In this work, we present NavGSim, a Gaussian Splatting-based simulator designed to generate high-fidelity, large-scale navigation environments. Built upon a hierarchical 3D Gaussian Splatting framework, NavGSim enables photorealistic rendering in expansive scenes spanning hundreds of square meters. To simulate navigation collisions, we introduce a Gaussian Splatting-based slice technique that directly extracts navigable areas from reconstructed Gaussians. Additionally, for ease of use, we provide comprehensive NavGSim APIs supporting multi-GPU development, including tools for custom scene reconstruction, robot configuration, policy training, and evaluation. To evaluate NavGSim's effectiveness, we train a Vision-Language-Action (VLA) model using trajectories collected from NavGSim and assess its performance in both simulated and real-world environments. Our results demonstrate that NavGSim significantly enhances the VLA model's scene understanding, enabling the policy to handle diverse navigation queries effectively.

Jiazhao Zhang, Kunyu Wang, Shaoan Wang et al.

A practical navigation agent must be capable of handling a wide range of interaction demands, such as following instructions, searching objects, answering questions, tracking people, and more. Existing models for embodied navigation fall short of serving as practical generalists in the real world, as they are often constrained by specific task configurations or pre-defined maps with discretized waypoints. In this work, we present Uni-NaVid, the first video-based vision-language-action (VLA) model designed to unify diverse embodied navigation tasks and enable seamless navigation for mixed long-horizon tasks in unseen real-world environments. Uni-NaVid achieves this by harmonizing the input and output data configurations for all commonly used embodied navigation tasks and thereby integrating all tasks in one model. For training Uni-NaVid, we collect 3.6 million navigation data samples in total from four essential navigation sub-tasks and foster synergy in learning across them. Extensive experiments on comprehensive navigation benchmarks clearly demonstrate the advantages of unification modeling in Uni-NaVid and show it achieves state-of-the-art performance. Additionally, real-world experiments confirm the model's effectiveness and efficiency, shedding light on its strong generalizability.

Mi Yan, Jiazhao Zhang, Yan Zhu et al.

Open-vocabulary 3D instance segmentation is cutting-edge for its ability to segment 3D instances without predefined categories. However, progress in 3D lags behind its 2D counterpart due to limited annotated 3D data. To address this, recent works first generate 2D open-vocabulary masks through 2D models and then merge them into 3D instances based on metrics calculated between two neighboring frames. In contrast to these local metrics, we propose a novel metric, view consensus rate, to enhance the utilization of multi-view observations. The key insight is that two 2D masks should be deemed part of the same 3D instance if a significant number of other 2D masks from different views contain both these two masks. Using this metric as edge weight, we construct a global mask graph where each mask is a node. Through iterative clustering of masks showing high view consensus, we generate a series of clusters, each representing a distinct 3D instance. Notably, our model is training-free. Through extensive experiments on publicly available datasets, including ScanNet++, ScanNet200 and MatterPort3D, we demonstrate that our method achieves state-of-the-art performance in open-vocabulary 3D instance segmentation. Our project page is at https://pku-epic.github.io/MaskClustering.

Zekun Qi, Wenyao Zhang, Yufei Ding et al. · pku, stanford

While spatial reasoning has made progress in object localization relationships, it often overlooks object orientation-a key factor in 6-DoF fine-grained manipulation. Traditional pose representations rely on pre-defined frames or templates, limiting generalization and semantic grounding. In this paper, we introduce the concept of semantic orientation, which defines object orientations using natural language in a reference-frame-free manner (e.g., the "plug-in" direction of a USB or the "handle" direction of a cup). To support this, we construct OrienText300K, a large-scale dataset of 3D objects annotated with semantic orientations, and develop PointSO, a general model for zero-shot semantic orientation prediction. By integrating semantic orientation into VLM agents, our SoFar framework enables 6-DoF spatial reasoning and generates robotic actions. Extensive experiments demonstrated the effectiveness and generalization of our SoFar, e.g., zero-shot 48.7% successful rate on Open6DOR and zero-shot 74.9% successful rate on SIMPLER-Env.

Wenyao Zhang, Hongsi Liu, Zekun Qi et al.

Recent advances in vision-language-action (VLA) models have shown promise in integrating image generation with action prediction to improve generalization and reasoning in robot manipulation. However, existing methods are limited to challenging image-based forecasting, which suffers from redundant information and lacks comprehensive and critical world knowledge, including dynamic, spatial and semantic information. To address these limitations, we propose DreamVLA, a novel VLA framework that integrates comprehensive world knowledge forecasting to enable inverse dynamics modeling, thereby establishing a perception-prediction-action loop for manipulation tasks. Specifically, DreamVLA introduces a dynamic-region-guided world knowledge prediction, integrated with the spatial and semantic cues, which provide compact yet comprehensive representations for action planning. This design aligns with how humans interact with the world by first forming abstract multimodal reasoning chains before acting. To mitigate interference among the dynamic, spatial and semantic information during training, we adopt a block-wise structured attention mechanism that masks their mutual attention, preventing information leakage and keeping each representation clean and disentangled. Moreover, to model the conditional distribution over future actions, we employ a diffusion-based transformer that disentangles action representations from shared latent features. Extensive experiments on both real-world and simulation environments demonstrate that DreamVLA achieves 76.7% success rate on real robot tasks and 4.44 average length on the CALVIN ABC-D benchmarks.

Shaoan Wang, Jiazhao Zhang, Minghan Li et al.

Embodied visual tracking is a fundamental skill in Embodied AI, enabling an agent to follow a specific target in dynamic environments using only egocentric vision. This task is inherently challenging as it requires both accurate target recognition and effective trajectory planning under conditions of severe occlusion and high scene dynamics. Existing approaches typically address this challenge through a modular separation of recognition and planning. In this work, we propose TrackVLA, a Vision-Language-Action (VLA) model that learns the synergy between object recognition and trajectory planning. Leveraging a shared LLM backbone, we employ a language modeling head for recognition and an anchor-based diffusion model for trajectory planning. To train TrackVLA, we construct an Embodied Visual Tracking Benchmark (EVT-Bench) and collect diverse difficulty levels of recognition samples, resulting in a dataset of 1.7 million samples. Through extensive experiments in both synthetic and real-world environments, TrackVLA demonstrates SOTA performance and strong generalizability. It significantly outperforms existing methods on public benchmarks in a zero-shot manner while remaining robust to high dynamics and occlusion in real-world scenarios at 10 FPS inference speed. Our project page is: https://pku-epic.github.io/TrackVLA-web.

Yihan Cao, Jiazhao Zhang, Zhinan Yu et al.

Object goal navigation (ObjectNav) is a fundamental task in embodied AI, requiring an agent to locate a target object in previously unseen environments. This task is particularly challenging because it requires both perceptual and cognitive processes, including object recognition and decision-making. While substantial advancements in perception have been driven by the rapid development of visual foundation models, progress on the cognitive aspect remains constrained, primarily limited to either implicit learning through simulator rollouts or explicit reliance on predefined heuristic rules. Inspired by neuroscientific findings demonstrating that humans maintain and dynamically update fine-grained cognitive states during object search tasks in novel environments, we propose CogNav, a framework designed to mimic this cognitive process using large language models. Specifically, we model the cognitive process using a finite state machine comprising fine-grained cognitive states, ranging from exploration to identification. Transitions between states are determined by a large language model based on a dynamically constructed heterogeneous cognitive map, which contains spatial and semantic information about the scene being explored. Extensive evaluations on the HM3D, MP3D, and RoboTHOR benchmarks demonstrate that our cognitive process modeling significantly improves the success rate of ObjectNav at least by relative 14% over the state-of-the-arts.

Chen Gao, Liankai Jin, Xingyu Peng et al.

Embodied navigation stands as a foundation pillar within the broader pursuit of embodied AI. However, previous navigation research is divided into different tasks/capabilities, e.g., ObjNav, ImgNav and VLN, where they differ in task objectives and modalities, making datasets and methods are designed individually. In this work, we take steps toward generalist navigation agents, which can follow free-form instructions that include arbitrary compounds of multi-modal and multi-capability. To achieve this, we propose a large-scale benchmark and corresponding method, termed OctoNav-Bench and OctoNav-R1. Specifically, OctoNav-Bench features continuous environments and is constructed via a designed annotation pipeline. We thoroughly craft instruction-trajectory pairs, where instructions are diverse in free-form with arbitrary modality and capability. Also, we construct a Think-Before-Action (TBA-CoT) dataset within OctoNav-Bench to provide the thinking process behind actions. For OctoNav-R1, we build it upon MLLMs and adapt it to a VLA-type model, which can produce low-level actions solely based on 2D visual observations. Moreover, we design a Hybrid Training Paradigm (HTP) that consists of three stages, i.e., Action-/TBA-SFT, Nav-GPRO, and Online RL stages. Each stage contains specifically designed learning policies and rewards. Importantly, for TBA-SFT and Nav-GRPO designs, we are inspired by the OpenAI-o1 and DeepSeek-R1, which show impressive reasoning ability via thinking-before-answer. Thus, we aim to investigate how to achieve thinking-before-action in the embodied navigation field, to improve model's reasoning ability toward generalists. Specifically, we propose TBA-SFT to utilize the TBA-CoT dataset to fine-tune the model as a cold-start phrase and then leverage Nav-GPRO to improve its thinking ability. Finally, OctoNav-R1 shows superior performance compared with previous methods.

Wenbo Cui, Chengyang Zhao, Songlin Wei et al. · berkeley

Effectively manipulating articulated objects in household scenarios is a crucial step toward achieving general embodied artificial intelligence. Mainstream research in 3D vision has primarily focused on manipulation through depth perception and pose detection. However, in real-world environments, these methods often face challenges due to imperfect depth perception, such as with transparent lids and reflective handles. Moreover, they generally lack the diversity in part-based interactions required for flexible and adaptable manipulation. To address these challenges, we introduced a large-scale part-centric dataset for articulated object manipulation that features both photo-realistic material randomization and detailed annotations of part-oriented, scene-level actionable interaction poses. We evaluated the effectiveness of our dataset by integrating it with several state-of-the-art methods for depth estimation and interaction pose prediction. Additionally, we proposed a novel modular framework that delivers superior and robust performance for generalizable articulated object manipulation. Our extensive experiments demonstrate that our dataset significantly improves the performance of depth perception and actionable interaction pose prediction in both simulation and real-world scenarios. More information and demos can be found at: https://pku-epic.github.io/GAPartManip/.

Yuhang Huang, Jiazhao Zhang, Shilong Zou et al.

Predictive manipulation has recently gained considerable attention in the Embodied AI community due to its potential to improve robot policy performance by leveraging predicted states. However, generating accurate future visual states of robot-object interactions from world models remains a well-known challenge, particularly in achieving high-quality pixel-level representations. To this end, we propose LaDi-WM, a world model that predicts the latent space of future states using diffusion modeling. Specifically, LaDi-WM leverages the well-established latent space aligned with pre-trained Visual Foundation Models (VFMs), which comprises both geometric features (DINO-based) and semantic features (CLIP-based). We find that predicting the evolution of the latent space is easier to learn and more generalizable than directly predicting pixel-level images. Building on LaDi-WM, we design a diffusion policy that iteratively refines output actions by incorporating forecasted states, thereby generating more consistent and accurate results. Extensive experiments on both synthetic and real-world benchmarks demonstrate that LaDi-WM significantly enhances policy performance by 27.9\% on the LIBERO-LONG benchmark and 20\% on the real-world scenario. Furthermore, our world model and policies achieve impressive generalizability in real-world experiments.

Yuqing Lan, Chenyang Zhu, Shuaifeng Zhi et al.

The introduction of the neural implicit representation has notably propelled the advancement of online dense reconstruction techniques. Compared to traditional explicit representations, such as TSDF, it improves the mapping completeness and memory efficiency. However, the lack of reconstruction details and the time-consuming learning of neural representations hinder the widespread application of neural-based methods to large-scale online reconstruction. We introduce RemixFusion, a novel residual-based mixed representation for scene reconstruction and camera pose estimation dedicated to high-quality and large-scale online RGB-D reconstruction. In particular, we propose a residual-based map representation comprised of an explicit coarse TSDF grid and an implicit neural module that produces residuals representing fine-grained details to be added to the coarse grid. Such mixed representation allows for detail-rich reconstruction with bounded time and memory budget, contrasting with the overly-smoothed results by the purely implicit representations, thus paving the way for high-quality camera tracking. Furthermore, we extend the residual-based representation to handle multi-frame joint pose optimization via bundle adjustment (BA). In contrast to the existing methods, which optimize poses directly, we opt to optimize pose changes. Combined with a novel technique for adaptive gradient amplification, our method attains better optimization convergence and global optimality. Furthermore, we adopt a local moving volume to factorize the mixed scene representation with a divide-and-conquer design to facilitate efficient online learning in our residual-based framework. Extensive experiments demonstrate that our method surpasses all state-of-the-art ones, including those based either on explicit or implicit representations, in terms of the accuracy of both mapping and tracking on large-scale scenes.

Jiahang Liu, Yunpeng Qi, Jiazhao Zhang et al.

Embodied Visual Tracking (EVT) is a fundamental ability that underpins practical applications, such as companion robots, guidance robots and service assistants, where continuously following moving targets is essential. Recent advances have enabled language-guided tracking in complex and unstructured scenes. However, existing approaches lack explicit spatial reasoning and effective temporal memory, causing failures under severe occlusions or in the presence of similar-looking distractors. To address these challenges, we present TrackVLA++, a novel Vision-Language-Action (VLA) model that enhances embodied visual tracking with two key modules, a spatial reasoning mechanism and a Target Identification Memory (TIM). The reasoning module introduces a Chain-of-Thought paradigm, termed Polar-CoT, which infers the target's relative position and encodes it as a compact polar-coordinate token for action prediction. Guided by these spatial priors, the TIM employs a gated update strategy to preserve long-horizon target memory, ensuring spatiotemporal consistency and mitigating target loss during extended occlusions. Extensive experiments show that TrackVLA++ achieves state-of-the-art performance on public benchmarks across both egocentric and multi-camera settings. On the challenging EVT-Bench DT split, TrackVLA++ surpasses the previous leading approach by 5.1 and 12, respectively. Furthermore, TrackVLA++ exhibits strong zero-shot generalization, enabling robust real-world tracking in dynamic and occluded scenarios.

Yijie Tang, Jiazhao Zhang, Yuqing Lan et al.

Online zero-shot 3D instance segmentation of a progressively reconstructed scene is both a critical and challenging task for embodied applications. With the success of visual foundation models (VFMs) in the image domain, leveraging 2D priors to address 3D online segmentation has become a prominent research focus. Since segmentation results provided by 2D priors often require spatial consistency to be lifted into final 3D segmentation, an efficient method for identifying spatial overlap among 2D masks is essential - yet existing methods rarely achieve this in real time, mainly limiting its use to offline approaches. To address this, we propose an efficient method that lifts 2D masks generated by VFMs into a unified 3D instance using a hashing technique. By employing voxel hashing for efficient 3D scene querying, our approach reduces the time complexity of costly spatial overlap queries from $O(n^2)$ to $O(n)$. Accurate spatial associations further enable 3D merging of 2D masks through simple similarity-based filtering in a zero-shot manner, making our approach more robust to incomplete and noisy data. Evaluated on the ScanNet and SceneNN benchmarks, our approach achieves state-of-the-art performance in online, zero-shot 3D instance segmentation with leading efficiency.

Yihan Cao, Jiazhao Zhang, Zhinan Yu et al.

Camera placement is crutial in multi-camera systems such as virtual reality, autonomous driving, and high-quality reconstruction. The camera placement challenge lies in the nonlinear nature of high-dimensional parameters and the unavailability of gradients for target functions like coverage and visibility. Consequently, most existing methods tackle this challenge by leveraging non-gradient-based optimization methods.In this work, we present a hybrid camera placement optimization approach that incorporates both gradient-based and non-gradient-based optimization methods. This design allows our method to enjoy the advantages of smooth optimization convergence and robustness from gradient-based and non-gradient-based optimization, respectively. To bridge the two disparate optimization methods, we propose a neural observation field, which implicitly encodes the coverage and observation quality. The neural observation field provides the measurements of the camera observations and corresponding gradients without the assumption of target scenes, making our method applicable to diverse scenarios, including 2D planar shapes, 3D objects, and room-scale 3D scenes.Extensive experiments on diverse datasets demonstrate that our method achieves state-of-the-art performance, while requiring only a fraction (8x less) of the typical computation time. Furthermore, we conducted a real-world experiment using a custom-built capture system, confirming the resilience of our approach to real-world environmental noise.

Anqi Li, Zhiyong Wang, Jiazhao Zhang et al.

Urban micromobility applications, such as delivery robots, demand reliable navigation across large-scale urban environments while following long-horizon route instructions. This task is particularly challenging due to the dynamic and unstructured nature of real-world city areas, yet most existing navigation methods remain tailored to short-scale and controllable scenarios. Effective urban micromobility requires two complementary levels of navigation skills: low-level capabilities such as point-goal reaching and obstacle avoidance, and high-level capabilities, such as route-visual alignment. To this end, we propose UrbanVLA, a route-conditioned Vision-Language-Action (VLA) framework designed for scalable urban navigation. Our method explicitly aligns noisy route waypoints with visual observations during execution, and subsequently plans trajectories to drive the robot. To enable UrbanVLA to master both levels of navigation, we employ a two-stage training pipeline. The process begins with Supervised Fine-Tuning (SFT) using simulated environments and trajectories parsed from web videos. This is followed by Reinforcement Fine-Tuning (RFT) on a mixture of simulation and real-world data, which enhances the model's safety and adaptability in real-world settings. Experiments demonstrate that UrbanVLA surpasses strong baselines by more than 55% in the SocialNav task on MetaUrban. Furthermore, UrbanVLA achieves reliable real-world navigation, showcasing both scalability to large-scale urban environments and robustness against real-world uncertainties.

Tianyu Xu, Jiawei Chen, Jiazhao Zhang et al.

Visual navigation policy is widely regarded as a promising direction, as it mimics humans by using egocentric visual observations for navigation. However, optical information of visual observations is difficult to be explicitly modeled like LiDAR point clouds or depth maps, which subsequently requires intelligent models and large-scale data. To this end, we propose to leverage the intelligence of the Vision-Language-Action (VLA) model to learn diverse navigation capabilities from synthetic expert data in a teacher-student manner. Specifically, we implement the VLA model, MM-Nav, as a multi-view VLA (with 360 observations) based on pretrained large language models and visual foundation models. For large-scale navigation data, we collect expert data from three reinforcement learning (RL) experts trained with privileged depth information in three challenging tailor-made environments for different navigation capabilities: reaching, squeezing, and avoiding. We iteratively train our VLA model using data collected online from RL experts, where the training ratio is dynamically balanced based on performance on individual capabilities. Through extensive experiments in synthetic environments, we demonstrate that our model achieves strong generalization capability. Moreover, we find that our student VLA model outperforms the RL teachers, demonstrating the synergistic effect of integrating multiple capabilities. Extensive real-world experiments further confirm the effectiveness of our method.

Yuqing Lan, Chenyang Zhu, Zhirui Gao et al.

Open-vocabulary 3D object detection has gained significant interest due to its critical applications in autonomous driving and embodied AI. Existing detection methods, whether offline or online, typically rely on dense point cloud reconstruction, which imposes substantial computational overhead and memory constraints, hindering real-time deployment in downstream tasks. To address this, we propose a novel reconstruction-free online framework tailored for memory-efficient and real-time 3D detection. Specifically, given streaming posed RGB-D video input, we leverage Cubify Anything as a pre-trained visual foundation model (VFM) for single-view 3D object detection by bounding boxes, coupled with CLIP to capture open-vocabulary semantics of detected objects. To fuse all detected bounding boxes across different views into a unified one, we employ an association module for correspondences of multi-views and an optimization module to fuse the 3D bounding boxes of the same instance predicted in multi-views. The association module utilizes 3D Non-Maximum Suppression (NMS) and a box correspondence matching module, while the optimization module uses an IoU-guided efficient random optimization technique based on particle filtering to enforce multi-view consistency of the 3D bounding boxes while minimizing computational complexity. Extensive experiments on ScanNetV2 and CA-1M datasets demonstrate that our method achieves state-of-the-art performance among online methods. Benefiting from this novel reconstruction-free paradigm for 3D object detection, our method exhibits great generalization abilities in various scenarios, enabling real-time perception even in environments exceeding 1000 square meters.

Jiazhao Zhang, Ying Hung, Chung-Ching Lin et al.

Choosing appropriate hyperparameters plays a crucial role in the success of neural networks as hyper-parameters directly control the behavior and performance of the training algorithms. To obtain efficient tuning, Bayesian optimization methods based on Gaussian process (GP) models are widely used. Despite numerous applications of Bayesian optimization in deep learning, the existing methodologies are developed based on a convenient but restrictive assumption that the tuning parameters are independent of each other. However, tuning parameters with conditional dependence are common in practice. In this paper, we focus on two types of them: branching and nested parameters. Nested parameters refer to those tuning parameters that exist only within a particular setting of another tuning parameter, and a parameter within which other parameters are nested is called a branching parameter. To capture the conditional dependence between branching and nested parameters, a unified Bayesian optimization framework is proposed. The sufficient conditions are rigorously derived to guarantee the validity of the kernel function, and the asymptotic convergence of the proposed optimization framework is proven under the continuum-armed-bandit setting. Based on the new GP model, which accounts for the dependent structure among input variables through a new kernel function, higher prediction accuracy and better optimization efficiency are observed in a series of synthetic simulations and real data applications of neural networks. Sensitivity analysis is also performed to provide insights into how changes in hyperparameter values affect prediction accuracy.

Jiazhao Zhang, Chenyang Zhu, Lintao Zheng et al.

Online reconstruction based on RGB-D sequences has thus far been restrained to relatively slow camera motions (<1m/s). Under very fast camera motion (e.g., 3m/s), the reconstruction can easily crumble even for the state-of-the-art methods. Fast motion brings two challenges to depth fusion: 1) the high nonlinearity of camera pose optimization due to large inter-frame rotations and 2) the lack of reliably trackable features due to motion blur. We propose to tackle the difficulties of fast-motion camera tracking in the absence of inertial measurements using random optimization, in particular, the Particle Filter Optimization (PFO). To surmount the computation-intensive particle sampling and update in standard PFO, we propose to accelerate the randomized search via updating a particle swarm template (PST). PST is a set of particles pre-sampled uniformly within the unit sphere in the 6D space of camera pose. Through moving and rescaling the pre-sampled PST guided by swarm intelligence, our method is able to drive tens of thousands of particles to locate and cover a good local optimum extremely fast and robustly. The particles, representing candidate poses, are evaluated with a fitness function defined based on depth-model conformance. Therefore, our method, being depth-only and correspondence-free, mitigates the motion blur impediment as ToF-based depths are often resilient to motion blur. Thanks to the efficient template-based particle set evolution and the effective fitness function, our method attains good quality pose tracking under fast camera motion (up to 4m/s) in a realtime framerate without including loop closure or global pose optimization. Through extensive evaluations on public datasets of RGB-D sequences, especially on a newly proposed benchmark of fast camera motion, we demonstrate the significant advantage of our method over the state of the arts.

Jiazhao Zhang, Chenyang Zhu, Lintao Zheng et al.

Online semantic 3D segmentation in company with real-time RGB-D reconstruction poses special challenges such as how to perform 3D convolution directly over the progressively fused 3D geometric data, and how to smartly fuse information from frame to frame. We propose a novel fusion-aware 3D point convolution which operates directly on the geometric surface being reconstructed and exploits effectively the inter-frame correlation for high quality 3D feature learning. This is enabled by a dedicated dynamic data structure which organizes the online acquired point cloud with global-local trees. Globally, we compile the online reconstructed 3D points into an incrementally growing coordinate interval tree, enabling fast point insertion and neighborhood query. Locally, we maintain the neighborhood information for each point using an octree whose construction benefits from the fast query of the global tree.Both levels of trees update dynamically and help the 3D convolution effectively exploits the temporal coherence for effective information fusion across RGB-D frames.

Lintao Zheng, Chenyang Zhu, Jiazhao Zhang et al.

We propose a novel approach to robot-operated active understanding of unknown indoor scenes, based on online RGBD reconstruction with semantic segmentation. In our method, the exploratory robot scanning is both driven by and targeting at the recognition and segmentation of semantic objects from the scene. Our algorithm is built on top of the volumetric depth fusion framework (e.g., KinectFusion) and performs real-time voxel-based semantic labeling over the online reconstructed volume. The robot is guided by an online estimated discrete viewing score field (VSF) parameterized over the 3D space of 2D location and azimuth rotation. VSF stores for each grid the score of the corresponding view, which measures how much it reduces the uncertainty (entropy) of both geometric reconstruction and semantic labeling. Based on VSF, we select the next best views (NBV) as the target for each time step. We then jointly optimize the traverse path and camera trajectory between two adjacent NBVs, through maximizing the integral viewing score (information gain) along path and trajectory. Through extensive evaluation, we show that our method achieves efficient and accurate online scene parsing during exploratory scanning.