Abstract:In multimodal large language models (MLLMs), the length of input visual tokens is often significantly greater than that of their textual counterparts, leading to a high inference cost. Many works aim to address this issue by removing redundant visual tokens. However, current approaches either rely on attention-based pruning, which retains numerous duplicate tokens, or use similarity-based pruning, overlooking the instruction relevance, consequently causing suboptimal performance. In this paper, we go beyond attention or similarity by proposing a novel visual token pruning method named CDPruner, which maximizes the conditional diversity of retained tokens. We first define the conditional similarity between visual tokens conditioned on the instruction, and then reformulate the token pruning problem with determinantal point process (DPP) to maximize the conditional diversity of the selected subset. The proposed CDPruner is training-free and model-agnostic, allowing easy application to various MLLMs. Extensive experiments across diverse MLLMs show that CDPruner establishes new state-of-the-art on various vision-language benchmarks. By maximizing conditional diversity through DPP, the selected subset better represents the input images while closely adhering to user instructions, thereby preserving strong performance even with high reduction ratios. When applied to LLaVA, CDPruner reduces FLOPs by 95\% and CUDA latency by 78\%, while maintaining 94\% of the original accuracy. Our code is available at https://github.com/Theia-4869/CDPruner.
Abstract:Video content comprehension is essential for various applications, ranging from video analysis to interactive systems. Despite advancements in large-scale vision-language models (VLMs), these models often struggle to capture the nuanced, spatiotemporal details essential for thorough video analysis. To address this gap, we introduce Video-CoT, a groundbreaking dataset designed to enhance spatiotemporal understanding using Chain-of-Thought (CoT) methodologies. Video-CoT contains 192,000 fine-grained spa-tiotemporal question-answer pairs and 23,000 high-quality CoT-annotated samples, providing a solid foundation for evaluating spatiotemporal understanding in video comprehension. Additionally, we provide a comprehensive benchmark for assessing these tasks, with each task featuring 750 images and tailored evaluation metrics. Our extensive experiments reveal that current VLMs face significant challenges in achieving satisfactory performance, high-lighting the difficulties of effective spatiotemporal understanding. Overall, the Video-CoT dataset and benchmark open new avenues for research in multimedia understanding and support future innovations in intelligent systems requiring advanced video analysis capabilities. By making these resources publicly available, we aim to encourage further exploration in this critical area. Project website:https://video-cot.github.io/ .
Abstract:Learning to control high-speed objects in the real world remains a challenging frontier in robotics. Table tennis serves as an ideal testbed for this problem, demanding both rapid interception of fast-moving balls and precise adjustment of their trajectories. This task presents two fundamental challenges: it requires a high-precision vision system capable of accurately predicting ball trajectories, and it necessitates intelligent strategic planning to ensure precise ball placement to target regions. The dynamic nature of table tennis, coupled with its real-time response requirements, makes it particularly well-suited for advancing robotic control capabilities in fast-paced, precision-critical domains. In this paper, we present SpikePingpong, a novel system that integrates spike-based vision with imitation learning for high-precision robotic table tennis. Our approach introduces two key attempts that directly address the aforementioned challenges: SONIC, a spike camera-based module that achieves millimeter-level precision in ball-racket contact prediction by compensating for real-world uncertainties such as air resistance and friction; and IMPACT, a strategic planning module that enables accurate ball placement to targeted table regions. The system harnesses a 20 kHz spike camera for high-temporal resolution ball tracking, combined with efficient neural network models for real-time trajectory correction and stroke planning. Experimental results demonstrate that SpikePingpong achieves a remarkable 91% success rate for 30 cm accuracy target area and 71% in the more challenging 20 cm accuracy task, surpassing previous state-of-the-art approaches by 38% and 37% respectively. These significant performance improvements enable the robust implementation of sophisticated tactical gameplay strategies, providing a new research perspective for robotic control in high-speed dynamic tasks.
Abstract:Spatial referring is a fundamental capability of embodied robots to interact with the 3D physical world. However, even with the powerful pretrained vision language models (VLMs), recent approaches are still not qualified to accurately understand the complex 3D scenes and dynamically reason about the instruction-indicated locations for interaction. To this end, we propose RoboRefer, a 3D-aware VLM that can first achieve precise spatial understanding by integrating a disentangled but dedicated depth encoder via supervised fine-tuning (SFT). Moreover, RoboRefer advances generalized multi-step spatial reasoning via reinforcement fine-tuning (RFT), with metric-sensitive process reward functions tailored for spatial referring tasks. To support SFT and RFT training, we introduce RefSpatial, a large-scale dataset of 20M QA pairs (2x prior), covering 31 spatial relations (vs. 15 prior) and supporting complex reasoning processes (up to 5 steps). In addition, we introduce RefSpatial-Bench, a challenging benchmark filling the gap in evaluating spatial referring with multi-step reasoning. Experiments show that SFT-trained RoboRefer achieves state-of-the-art spatial understanding, with an average success rate of 89.6%. RFT-trained RoboRefer further outperforms all other baselines by a large margin, even surpassing Gemini-2.5-Pro by 17.4% in average accuracy on RefSpatial-Bench. Notably, RoboRefer can be integrated with various control policies to execute long-horizon, dynamic tasks across diverse robots (e,g., UR5, G1 humanoid) in cluttered real-world scenes.
Abstract:Recent advancements in world models have revolutionized dynamic environment simulation, allowing systems to foresee future states and assess potential actions. In autonomous driving, these capabilities help vehicles anticipate the behavior of other road users, perform risk-aware planning, accelerate training in simulation, and adapt to novel scenarios, thereby enhancing safety and reliability. Current approaches exhibit deficiencies in maintaining robust 3D geometric consistency or accumulating artifacts during occlusion handling, both critical for reliable safety assessment in autonomous navigation tasks. To address this, we introduce GeoDrive, which explicitly integrates robust 3D geometry conditions into driving world models to enhance spatial understanding and action controllability. Specifically, we first extract a 3D representation from the input frame and then obtain its 2D rendering based on the user-specified ego-car trajectory. To enable dynamic modeling, we propose a dynamic editing module during training to enhance the renderings by editing the positions of the vehicles. Extensive experiments demonstrate that our method significantly outperforms existing models in both action accuracy and 3D spatial awareness, leading to more realistic, adaptable, and reliable scene modeling for safer autonomous driving. Additionally, our model can generalize to novel trajectories and offers interactive scene editing capabilities, such as object editing and object trajectory control.
Abstract:We propose a novel framework for comprehensive indoor 3D reconstruction using Gaussian representations, called OmniIndoor3D. This framework enables accurate appearance, geometry, and panoptic reconstruction of diverse indoor scenes captured by a consumer-level RGB-D camera. Since 3DGS is primarily optimized for photorealistic rendering, it lacks the precise geometry critical for high-quality panoptic reconstruction. Therefore, OmniIndoor3D first combines multiple RGB-D images to create a coarse 3D reconstruction, which is then used to initialize the 3D Gaussians and guide the 3DGS training. To decouple the optimization conflict between appearance and geometry, we introduce a lightweight MLP that adjusts the geometric properties of 3D Gaussians. The introduced lightweight MLP serves as a low-pass filter for geometry reconstruction and significantly reduces noise in indoor scenes. To improve the distribution of Gaussian primitives, we propose a densification strategy guided by panoptic priors to encourage smoothness on planar surfaces. Through the joint optimization of appearance, geometry, and panoptic reconstruction, OmniIndoor3D provides comprehensive 3D indoor scene understanding, which facilitates accurate and robust robotic navigation. We perform thorough evaluations across multiple datasets, and OmniIndoor3D achieves state-of-the-art results in appearance, geometry, and panoptic reconstruction. We believe our work bridges a critical gap in indoor 3D reconstruction. The code will be released at: https://ucwxb.github.io/OmniIndoor3D/
Abstract:Neuromorphic Visual Systems, such as spike cameras, have attracted considerable attention due to their ability to capture clear textures under dynamic conditions. This capability effectively mitigates issues related to motion and aperture blur. However, in contrast to conventional RGB modalities that provide dense spatial information, these systems generate binary, spatially sparse frames as a trade-off for temporally rich visual streams. In this context, generative models emerge as a promising solution to address the inherent limitations of sparse data. These models not only facilitate the conditional fusion of existing information from both spike and RGB modalities but also enable the conditional generation based on latent priors. In this study, we introduce a robust generative processing framework named SpikeGen, designed for visual spike streams captured by spike cameras. We evaluate this framework across multiple tasks involving mixed spike-RGB modalities, including conditional image/video deblurring, dense frame reconstruction from spike streams, and high-speed scene novel-view synthesis. Supported by comprehensive experimental results, we demonstrate that leveraging the latent space operation abilities of generative models allows us to effectively address the sparsity of spatial information while fully exploiting the temporal richness of spike streams, thereby promoting a synergistic enhancement of different visual modalities.
Abstract:The development of artificial intelligence demands that models incrementally update knowledge by Continual Learning (CL) to adapt to open-world environments. To meet privacy and security requirements, Continual Unlearning (CU) emerges as an important problem, aiming to sequentially forget particular knowledge acquired during the CL phase. However, existing unlearning methods primarily focus on single-shot joint forgetting and face significant limitations when applied to CU. First, most existing methods require access to the retained dataset for re-training or fine-tuning, violating the inherent constraint in CL that historical data cannot be revisited. Second, these methods often suffer from a poor trade-off between system efficiency and model fidelity, making them vulnerable to being overwhelmed or degraded by adversaries through deliberately frequent requests. In this paper, we identify that the limitations of existing unlearning methods stem fundamentally from their reliance on gradient-based updates. To bridge the research gap at its root, we propose a novel gradient-free method for CU, named Analytic Continual Unlearning (ACU), for efficient and exact forgetting with historical data privacy preservation. In response to each unlearning request, our ACU recursively derives an analytical (i.e., closed-form) solution in an interpretable manner using the least squares method. Theoretical and experimental evaluations validate the superiority of our ACU on unlearning effectiveness, model fidelity, and system efficiency.
Abstract:Federated Continual Learning (FCL) enables distributed clients to collaboratively train a global model from online task streams in dynamic real-world scenarios. However, existing FCL methods face challenges of both spatial data heterogeneity among distributed clients and temporal data heterogeneity across online tasks. Such data heterogeneity significantly degrades the model performance with severe spatial-temporal catastrophic forgetting of local and past knowledge. In this paper, we identify that the root cause of this issue lies in the inherent vulnerability and sensitivity of gradients to non-IID data. To fundamentally address this issue, we propose a gradient-free method, named Analytic Federated Continual Learning (AFCL), by deriving analytical (i.e., closed-form) solutions from frozen extracted features. In local training, our AFCL enables single-epoch learning with only a lightweight forward-propagation process for each client. In global aggregation, the server can recursively and efficiently update the global model with single-round aggregation. Theoretical analyses validate that our AFCL achieves spatio-temporal invariance of non-IID data. This ideal property implies that, regardless of how heterogeneous the data are distributed across local clients and online tasks, the aggregated model of our AFCL remains invariant and identical to that of centralized joint learning. Extensive experiments show the consistent superiority of our AFCL over state-of-the-art baselines across various benchmark datasets and settings.
Abstract:Large-scale pre-training using videos has proven effective for robot learning. However, the models pre-trained on such data can be suboptimal for robot learning due to the significant visual gap between human hands and those of different robots. To remedy this, we propose H2R, a simple data augmentation technique that detects human hand keypoints, synthesizes robot motions in simulation, and composites rendered robots into egocentric videos. This process explicitly bridges the visual gap between human and robot embodiments during pre-training. We apply H2R to augment large-scale egocentric human video datasets such as Ego4D and SSv2, replacing human hands with simulated robotic arms to generate robot-centric training data. Based on this, we construct and release a family of 1M-scale datasets covering multiple robot embodiments (UR5 with gripper/Leaphand, Franka) and data sources (SSv2, Ego4D). To verify the effectiveness of the augmentation pipeline, we introduce a CLIP-based image-text similarity metric that quantitatively evaluates the semantic fidelity of robot-rendered frames to the original human actions. We validate H2R across three simulation benchmarks: Robomimic, RLBench and PushT and real-world manipulation tasks with a UR5 robot equipped with Gripper and Leaphand end-effectors. H2R consistently improves downstream success rates, yielding gains of 5.0%-10.2% in simulation and 6.7%-23.3% in real-world tasks across various visual encoders and policy learning methods. These results indicate that H2R improves the generalization ability of robotic policies by mitigating the visual discrepancies between human and robot domains.