EvoMoE: Expert Evolution in Mixture of Experts for Multimodal Large Language Models

Recent advancements have shown that the Mixture of Experts (MoE) approach significantly enhances the capacity of large language models (LLMs) and improves performance on downstream tasks. Building on these promising results, multi-modal large language models (MLLMs) have increasingly adopted MoE techniques. However, existing multi-modal MoE tuning methods typically face two key challenges: expert uniformity and router rigidity. Expert uniformity occurs because MoE experts are often initialized by simply replicating the FFN parameters from LLMs, leading to homogenized expert functions and weakening the intended diversification of the MoE architecture. Meanwhile, router rigidity stems from the prevalent use of static linear routers for expert selection, which fail to distinguish between visual and textual tokens, resulting in similar expert distributions for image and text. To address these limitations, we propose EvoMoE, an innovative MoE tuning framework. EvoMoE introduces a meticulously designed expert initialization strategy that progressively evolves multiple robust experts from a single trainable expert, a process termed expert evolution that specifically targets severe expert homogenization. Furthermore, we introduce the Dynamic Token-aware Router (DTR), a novel routing mechanism that allocates input tokens to appropriate experts based on their modality and intrinsic token values. This dynamic routing is facilitated by hypernetworks, which dynamically generate routing weights tailored for each individual token. Extensive experiments demonstrate that EvoMoE significantly outperforms other sparse MLLMs across a variety of multi-modal benchmarks, including MME, MMBench, TextVQA, and POPE. Our results highlight the effectiveness of EvoMoE in enhancing the performance of MLLMs by addressing the critical issues of expert uniformity and router rigidity.

Cross from Left to Right Brain: Adaptive Text Dreamer for Vision-and-Language Navigation

Vision-and-Language Navigation (VLN) requires the agent to navigate by following natural instructions under partial observability, making it difficult to align perception with language. Recent methods mitigate this by imagining future scenes, yet they rely on vision-based synthesis, leading to high computational cost and redundant details. To this end, we propose to adaptively imagine key environmental semantics via \textit{language} form, enabling a more reliable and efficient strategy. Specifically, we introduce a novel Adaptive Text Dreamer (ATD), a dual-branch self-guided imagination policy built upon a large language model (LLM). ATD is designed with a human-like left-right brain architecture, where the left brain focuses on logical integration, and the right brain is responsible for imaginative prediction of future scenes. To achieve this, we fine-tune only the Q-former within both brains to efficiently activate domain-specific knowledge in the LLM, enabling dynamic updates of logical reasoning and imagination during navigation. Furthermore, we introduce a cross-interaction mechanism to regularize the imagined outputs and inject them into a navigation expert module, allowing ATD to jointly exploit both the reasoning capacity of the LLM and the expertise of the navigation model. We conduct extensive experiments on the R2R benchmark, where ATD achieves state-of-the-art performance with fewer parameters. The code is \href{https://github.com/zhangpingrui/Adaptive-Text-Dreamer}{here}.

AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations

Visual grounding (VG) aims to localize target objects in an image based on natural language descriptions. In this paper, we propose AerialVG, a new task focusing on visual grounding from aerial views. Compared to traditional VG, AerialVG poses new challenges, \emph{e.g.}, appearance-based grounding is insufficient to distinguish among multiple visually similar objects, and positional relations should be emphasized. Besides, existing VG models struggle when applied to aerial imagery, where high-resolution images cause significant difficulties. To address these challenges, we introduce the first AerialVG dataset, consisting of 5K real-world aerial images, 50K manually annotated descriptions, and 103K objects. Particularly, each annotation in AerialVG dataset contains multiple target objects annotated with relative spatial relations, requiring models to perform comprehensive spatial reasoning. Furthermore, we propose an innovative model especially for the AerialVG task, where a Hierarchical Cross-Attention is devised to focus on target regions, and a Relation-Aware Grounding module is designed to infer positional relations. Experimental results validate the effectiveness of our dataset and method, highlighting the importance of spatial reasoning in aerial visual grounding. The code and dataset will be released.

Think Small, Act Big: Primitive Prompt Learning for Lifelong Robot Manipulation

Building a lifelong robot that can effectively leverage prior knowledge for continuous skill acquisition remains significantly challenging. Despite the success of experience replay and parameter-efficient methods in alleviating catastrophic forgetting problem, naively applying these methods causes a failure to leverage the shared primitives between skills. To tackle these issues, we propose Primitive Prompt Learning (PPL), to achieve lifelong robot manipulation via reusable and extensible primitives. Within our two stage learning scheme, we first learn a set of primitive prompts to represent shared primitives through multi-skills pre-training stage, where motion-aware prompts are learned to capture semantic and motion shared primitives across different skills. Secondly, when acquiring new skills in lifelong span, new prompts are appended and optimized with frozen pretrained prompts, boosting the learning via knowledge transfer from old skills to new ones. For evaluation, we construct a large-scale skill dataset and conduct extensive experiments in both simulation and real-world tasks, demonstrating PPL's superior performance over state-of-the-art methods.

MoMa-Kitchen: A 100K+ Benchmark for Affordance-Grounded Last-Mile Navigation in Mobile Manipulation

In mobile manipulation, navigation and manipulation are often treated as separate problems, resulting in a significant gap between merely approaching an object and engaging with it effectively. Many navigation approaches primarily define success by proximity to the target, often overlooking the necessity for optimal positioning that facilitates subsequent manipulation. To address this, we introduce MoMa-Kitchen, a benchmark dataset comprising over 100k samples that provide training data for models to learn optimal final navigation positions for seamless transition to manipulation. Our dataset includes affordance-grounded floor labels collected from diverse kitchen environments, in which robotic mobile manipulators of different models attempt to grasp target objects amidst clutter. Using a fully automated pipeline, we simulate diverse real-world scenarios and generate affordance labels for optimal manipulation positions. Visual data are collected from RGB-D inputs captured by a first-person view camera mounted on the robotic arm, ensuring consistency in viewpoint during data collection. We also develop a lightweight baseline model, NavAff, for navigation affordance grounding that demonstrates promising performance on the MoMa-Kitchen benchmark. Our approach enables models to learn affordance-based final positioning that accommodates different arm types and platform heights, thereby paving the way for more robust and generalizable integration of navigation and manipulation in embodied AI. Project page: \href{https://momakitchen.github.io/}{https://momakitchen.github.io/}.

Multi-Grained Feature Pruning for Video-Based Human Pose Estimation

Human pose estimation, with its broad applications in action recognition and motion capture, has experienced significant advancements. However, current Transformer-based methods for video pose estimation often face challenges in managing redundant temporal information and achieving fine-grained perception because they only focus on processing low-resolution features. To address these challenges, we propose a novel multi-scale resolution framework that encodes spatio-temporal representations at varying granularities and executes fine-grained perception compensation. Furthermore, we employ a density peaks clustering method to dynamically identify and prioritize tokens that offer important semantic information. This strategy effectively prunes redundant feature tokens, especially those arising from multi-frame features, thereby optimizing computational efficiency without sacrificing semantic richness. Empirically, it sets new benchmarks for both performance and efficiency on three large-scale datasets. Our method achieves a 93.8% improvement in inference speed compared to the baseline, while also enhancing pose estimation accuracy, reaching 87.4 mAP on the PoseTrack2017 dataset.

OpenFly: A Versatile Toolchain and Large-scale Benchmark for Aerial Vision-Language Navigation

Vision-Language Navigation (VLN) aims to guide agents through an environment by leveraging both language instructions and visual cues, playing a pivotal role in embodied AI. Indoor VLN has been extensively studied, whereas outdoor aerial VLN remains underexplored. The potential reason is that outdoor aerial view encompasses vast areas, making data collection more challenging, which results in a lack of benchmarks. To address this problem, we propose OpenFly, a platform comprising a versatile toolchain and large-scale benchmark for aerial VLN. Firstly, we develop a highly automated toolchain for data collection, enabling automatic point cloud acquisition, scene semantic segmentation, flight trajectory creation, and instruction generation. Secondly, based on the toolchain, we construct a large-scale aerial VLN dataset with 100k trajectories, covering diverse heights and lengths across 18 scenes. The corresponding visual data are generated using various rendering engines and advanced techniques, including Unreal Engine, GTA V, Google Earth, and 3D Gaussian Splatting (3D GS). All data exhibit high visual quality. Particularly, 3D GS supports real-to-sim rendering, further enhancing the realism of the dataset. Thirdly, we propose OpenFly-Agent, a keyframe-aware VLN model, which takes language instructions, current observations, and historical keyframes as input, and outputs flight actions directly. Extensive analyses and experiments are conducted, showcasing the superiority of our OpenFly platform and OpenFly-Agent. The toolchain, dataset, and codes will be open-sourced.

Exploring the Potential of Encoder-free Architectures in 3D LMMs

Encoder-free architectures have been preliminarily explored in the 2D visual domain, yet it remains an open question whether they can be effectively applied to 3D understanding scenarios. In this paper, we present the first comprehensive investigation into the potential of encoder-free architectures to overcome the challenges of encoder-based 3D Large Multimodal Models (LMMs). These challenges include the failure to adapt to varying point cloud resolutions and the point features from the encoder not meeting the semantic needs of Large Language Models (LLMs). We identify key aspects for 3D LMMs to remove the encoder and enable the LLM to assume the role of the 3D encoder: 1) We propose the LLM-embedded Semantic Encoding strategy in the pre-training stage, exploring the effects of various point cloud self-supervised losses. And we present the Hybrid Semantic Loss to extract high-level semantics. 2) We introduce the Hierarchical Geometry Aggregation strategy in the instruction tuning stage. This incorporates inductive bias into the LLM early layers to focus on the local details of the point clouds. To the end, we present the first Encoder-free 3D LMM, ENEL. Our 7B model rivals the current state-of-the-art model, ShapeLLM-13B, achieving 55.0%, 50.92%, and 42.7% on the classification, captioning, and VQA tasks, respectively. Our results demonstrate that the encoder-free architecture is highly promising for replacing encoder-based architectures in the field of 3D understanding. The code is released at https://github.com/Ivan-Tang-3D/ENEL

SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model

In this paper, we claim that spatial understanding is the keypoint in robot manipulation, and propose SpatialVLA to explore effective spatial representations for the robot foundation model. Specifically, we introduce Ego3D Position Encoding to inject 3D information into the input observations of the visual-language-action model, and propose Adaptive Action Grids to represent spatial robot movement actions with adaptive discretized action grids, facilitating learning generalizable and transferrable spatial action knowledge for cross-robot control. SpatialVLA is first pre-trained on top of a vision-language model with 1.1 Million real-world robot episodes, to learn a generalist manipulation policy across multiple robot environments and tasks. After pre-training, SpatialVLA is directly applied to perform numerous tasks in a zero-shot manner. The superior results in both simulation and real-world robots demonstrate its advantage of inferring complex robot motion trajectories and its strong in-domain multi-task generalization ability. We further show the proposed Adaptive Action Grids offer a new and effective way to fine-tune the pre-trained SpatialVLA model for new simulation and real-world setups, where the pre-learned action grids are re-discretized to capture robot-specific spatial action movements of new setups. The superior results from extensive evaluations demonstrate the exceptional in-distribution generalization and out-of-distribution adaptation capability, highlighting the crucial benefit of the proposed spatial-aware representations for generalist robot policy learning. All the details and codes will be open-sourced.

SpatioTemporal Learning for Human Pose Estimation in Sparsely-Labeled Videos

Human pose estimation in videos remains a challenge, largely due to the reliance on extensive manual annotation of large datasets, which is expensive and labor-intensive. Furthermore, existing approaches often struggle to capture long-range temporal dependencies and overlook the complementary relationship between temporal pose heatmaps and visual features. To address these limitations, we introduce STDPose, a novel framework that enhances human pose estimation by learning spatiotemporal dynamics in sparsely-labeled videos. STDPose incorporates two key innovations: 1) A novel Dynamic-Aware Mask to capture long-range motion context, allowing for a nuanced understanding of pose changes. 2) A system for encoding and aggregating spatiotemporal representations and motion dynamics to effectively model spatiotemporal relationships, improving the accuracy and robustness of pose estimation. STDPose establishes a new performance benchmark for both video pose propagation (i.e., propagating pose annotations from labeled frames to unlabeled frames) and pose estimation tasks, across three large-scale evaluation datasets. Additionally, utilizing pseudo-labels generated by pose propagation, STDPose achieves competitive performance with only 26.7% labeled data.