Feifei Feng

Junjie Wen, Yichen Zhu, Jinming Li et al.

Vision-Language-Action (VLA) models have shown remarkable potential in visuomotor control and instruction comprehension through end-to-end learning processes. However, current VLA models face significant challenges: they are slow during inference and require extensive pre-training on large amounts of robotic data, making real-world deployment difficult. In this paper, we introduce a new family of compact vision-language-action models, called TinyVLA, which offers two key advantages over existing VLA models: (1) faster inference speeds, and (2) improved data efficiency, eliminating the need for pre-training stage. Our framework incorporates two essential components to build TinyVLA: (1) initializing the policy backbone with robust, high-speed multimodal models, and (2) integrating a diffusion policy decoder during fine-tuning to enable precise robot actions. We conducted extensive evaluations of TinyVLA in both simulation and on real robots, demonstrating that our approach significantly outperforms the state-of-the-art VLA model, OpenVLA, in terms of speed and data efficiency, while delivering comparable or superior performance. Additionally, TinyVLA exhibits strong generalization capabilities across various dimensions, including language instructions, novel objects, unseen positions, changes in object appearance, background variations, and environmental shifts, often matching or exceeding the performance of OpenVLA. We believe that \methodname offers an interesting perspective on utilizing pre-trained multimodal models for policy learning. Our project is at https://tiny-vla.github.io.

Haowen Wang, Mingyuan Wang, Zhengping Che et al.

The raw depth image captured by the indoor depth sensor usually has an extensive range of missing depth values due to inherent limitations such as the inability to perceive transparent objects and limited distance range. The incomplete depth map burdens many downstream vision tasks, and a rising number of depth completion methods have been proposed to alleviate this issue. While most existing methods can generate accurate dense depth maps from sparse and uniformly sampled depth maps, they are not suitable for complementing the large contiguous regions of missing depth values, which is common and critical. In this paper, we design a novel two-branch end-to-end fusion network, which takes a pair of RGB and incomplete depth images as input to predict a dense and completed depth map. The first branch employs an encoder-decoder structure to regress the local dense depth values from the raw depth map, with the help of local guidance information extracted from the RGB image. In the other branch, we propose an RGB-depth fusion GAN to transfer the RGB image to the fine-grained textured depth map. We adopt adaptive fusion modules named W-AdaIN to propagate the features across the two branches, and we append a confidence fusion head to fuse the two outputs of the branches for the final depth map. Extensive experiments on NYU-Depth V2 and SUN RGB-D demonstrate that our proposed method clearly improves the depth completion performance, especially in a more realistic setting of indoor environments with the help of the pseudo depth map.

Yaomin Huang, Ning Liu, Zhengping Che et al.

Channel pruning can effectively reduce both computational cost and memory footprint of the original network while keeping a comparable accuracy performance. Though great success has been achieved in channel pruning for 2D image-based convolutional networks (CNNs), existing works seldom extend the channel pruning methods to 3D point-based neural networks (PNNs). Directly implementing the 2D CNN channel pruning methods to PNNs undermine the performance of PNNs because of the different representations of 2D images and 3D point clouds as well as the network architecture disparity. In this paper, we proposed CP$^3$, which is a Channel Pruning Plug-in for Point-based network. CP$^3$ is elaborately designed to leverage the characteristics of point clouds and PNNs in order to enable 2D channel pruning methods for PNNs. Specifically, it presents a coordinate-enhanced channel importance metric to reflect the correlation between dimensional information and individual channel features, and it recycles the discarded points in PNN's sampling process and reconsiders their potentially-exclusive information to enhance the robustness of channel pruning. Experiments on various PNN architectures show that CP$^3$ constantly improves state-of-the-art 2D CNN pruning approaches on different point cloud tasks. For instance, our compressed PointNeXt-S on ScanObjectNN achieves an accuracy of 88.52% with a pruning rate of 57.8%, outperforming the baseline pruning methods with an accuracy gain of 1.94%.

Yaomin Huang, Xinmei Liu, Yichen Zhu et al.

Detecting 3D objects from point clouds is a practical yet challenging task that has attracted increasing attention recently. In this paper, we propose a Label-Guided auxiliary training method for 3D object detection (LG3D), which serves as an auxiliary network to enhance the feature learning of existing 3D object detectors. Specifically, we propose two novel modules: a Label-Annotation-Inducer that maps annotations and point clouds in bounding boxes to task-specific representations and a Label-Knowledge-Mapper that assists the original features to obtain detection-critical representations. The proposed auxiliary network is discarded in inference and thus has no extra computational cost at test time. We conduct extensive experiments on both indoor and outdoor datasets to verify the effectiveness of our approach. For example, our proposed LG3D improves VoteNet by 2.5% and 3.1% mAP on the SUN RGB-D and ScanNetV2 datasets, respectively.

Haowen Wang, Zhengping Che, Yufan Yang et al.

Raw depth images captured in indoor scenarios frequently exhibit extensive missing values due to the inherent limitations of the sensors and environments. For example, transparent materials frequently elude detection by depth sensors; surfaces may introduce measurement inaccuracies due to their polished textures, extended distances, and oblique incidence angles from the sensor. The presence of incomplete depth maps imposes significant challenges for subsequent vision applications, prompting the development of numerous depth completion techniques to mitigate this problem. Numerous methods excel at reconstructing dense depth maps from sparse samples, but they often falter when faced with extensive contiguous regions of missing depth values, a prevalent and critical challenge in indoor environments. To overcome these challenges, we design a novel two-branch end-to-end fusion network named RDFC-GAN, which takes a pair of RGB and incomplete depth images as input to predict a dense and completed depth map. The first branch employs an encoder-decoder structure, by adhering to the Manhattan world assumption and utilizing normal maps from RGB-D information as guidance, to regress the local dense depth values from the raw depth map. The other branch applies an RGB-depth fusion CycleGAN, adept at translating RGB imagery into detailed, textured depth maps while ensuring high fidelity through cycle consistency. We fuse the two branches via adaptive fusion modules named W-AdaIN and train the model with the help of pseudo depth maps. Comprehensive evaluations on NYU-Depth V2 and SUN RGB-D datasets show that our method significantly enhances depth completion performance particularly in realistic indoor settings.

Haowen Wang, Zhipeng Fan, Zhen Zhao et al.

Estimating 6D poses and reconstructing 3D shapes of objects in open-world scenes from RGB-depth image pairs is challenging. Many existing methods rely on learning geometric features that correspond to specific templates while disregarding shape variations and pose differences among objects in the same category. As a result, these methods underperform when handling unseen object instances in complex environments. In contrast, other approaches aim to achieve category-level estimation and reconstruction by leveraging normalized geometric structure priors, but the static prior-based reconstruction struggles with substantial intra-class variations. To solve these problems, we propose the DTF-Net, a novel framework for pose estimation and shape reconstruction based on implicit neural fields of object categories. In DTF-Net, we design a deformable template field to represent the general category-wise shape latent features and intra-category geometric deformation features. The field establishes continuous shape correspondences, deforming the category template into arbitrary observed instances to accomplish shape reconstruction. We introduce a pose regression module that shares the deformation features and template codes from the fields to estimate the accurate 6D pose of each object in the scene. We integrate a multi-modal representation extraction module to extract object features and semantic masks, enabling end-to-end inference. Moreover, during training, we implement a shape-invariant training strategy and a viewpoint sampling method to further enhance the model's capability to extract object pose features. Extensive experiments on the REAL275 and CAMERA25 datasets demonstrate the superiority of DTF-Net in both synthetic and real scenes. Furthermore, we show that DTF-Net effectively supports grasping tasks with a real robot arm.

Mingze Wei, Yaomin Huang, Zhiyuan Xu et al.

In this paper, we propose a novel representation for grasping using contacts between multi-finger robotic hands and objects to be manipulated. This representation significantly reduces the prediction dimensions and accelerates the learning process. We present an effective end-to-end network, CMG-Net, for grasping unknown objects in a cluttered environment by efficiently predicting multi-finger grasp poses and hand configurations from a single-shot point cloud. Moreover, we create a synthetic grasp dataset that consists of five thousand cluttered scenes, 80 object categories, and 20 million annotations. We perform a comprehensive empirical study and demonstrate the effectiveness of our grasping representation and CMG-Net. Our work significantly outperforms the state-of-the-art for three-finger robotic hands. We also demonstrate that the model trained using synthetic data performs very well for real robots.

Minjie Zhu, Yichen Zhu, Xin Liu et al.

Multimodal Large Language Models (MLLMs) have showcased impressive skills in tasks related to visual understanding and reasoning. Yet, their widespread application faces obstacles due to the high computational demands during both the training and inference phases, restricting their use to a limited audience within the research and user communities. In this paper, we investigate the design aspects of Multimodal Small Language Models (MSLMs) and propose an efficient multimodal assistant named Mipha, which is designed to create synergy among various aspects: visual representation, language models, and optimization strategies. We show that without increasing the volume of training data, our Mipha-3B outperforms the state-of-the-art large MLLMs, especially LLaVA-1.5-13B, on multiple benchmarks. Through detailed discussion, we provide insights and guidelines for developing strong MSLMs that rival the capabilities of MLLMs. Our code is available at https://github.com/zhuyiche/llava-phi.

Xinwang Chen, Ning Liu, Yichen Zhu et al.

Transformer-based Diffusion Probabilistic Models (DPMs) have shown more potential than CNN-based DPMs, yet their extensive computational requirements hinder widespread practical applications. To reduce the computation budget of transformer-based DPMs, this work proposes the Efficient Diffusion Transformer (EDT) framework. The framework includes a lightweight-design diffusion model architecture, and a training-free Attention Modulation Matrix and its alternation arrangement in EDT inspired by human-like sketching. Additionally, we propose a token relation-enhanced masking training strategy tailored explicitly for EDT to augment its token relation learning capability. Our extensive experiments demonstrate the efficacy of EDT. The EDT framework reduces training and inference costs and surpasses existing transformer-based diffusion models in image synthesis performance, thereby achieving a significant overall enhancement. With lower FID, EDT-S, EDT-B, and EDT-XL attained speed-ups of 3.93x, 2.84x, and 1.92x respectively in the training phase, and 2.29x, 2.29x, and 2.22x respectively in inference, compared to the corresponding sizes of MDTv2. The source code is released at https://github.com/xinwangChen/EDT.

Jinming Li, Yichen Zhu, Zhiyuan Xu et al.

It is fundamentally challenging for robots to serve as useful assistants in human environments because this requires addressing a spectrum of sub-problems across robotics, including perception, language understanding, reasoning, and planning. The recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated their exceptional abilities in solving complex mathematical problems, mastering commonsense and abstract reasoning. This has led to the recent utilization of MLLMs as the brain in robotic systems, enabling these models to conduct high-level planning prior to triggering low-level control actions for task execution. However, it remains uncertain whether existing MLLMs are reliable in serving the brain role of robots. In this study, we introduce the first benchmark for evaluating Multimodal LLM for Robotic (MMRo) benchmark, which tests the capability of MLLMs for robot applications. Specifically, we identify four essential capabilities perception, task planning, visual reasoning, and safety measurement that MLLMs must possess to qualify as the robot's central processing unit. We have developed several scenarios for each capability, resulting in a total of 14 metrics for evaluation. We present experimental results for various MLLMs, including both commercial and open-source models, to assess the performance of existing systems. Our findings indicate that no single model excels in all areas, suggesting that current MLLMs are not yet trustworthy enough to serve as the cognitive core for robots. Our data can be found in https://mm-robobench.github.io/.

Junjie Wen, Yichen Zhu, Jinming Li et al.

Enabling robots to perform diverse tasks across varied environments is a central challenge in robot learning. While vision-language-action (VLA) models have shown promise for generalizable robot skills, realizing their full potential requires addressing limitations in action representation and efficient training. Current VLA models often focus on scaling the vision-language model (VLM) component, while the action space representation remains a critical bottleneck. This paper introduces DexVLA, a novel framework designed to enhance the efficiency and generalization capabilities of VLAs for complex, long-horizon tasks across diverse robot embodiments. DexVLA features a novel diffusion-based action expert, scaled to one billion parameters, designed for cross-embodiment learning. A novel embodiment curriculum learning strategy facilitates efficient training: (1) pre-training the diffusion expert that is separable from the VLA on cross-embodiment data, (2) aligning the VLA model to specific embodiments, and (3) post-training for rapid adaptation to new tasks. We conduct comprehensive experiments across multiple embodiments, including single-arm, bimanual, and dexterous hand, demonstrating DexVLA's adaptability to challenging tasks without task-specific adaptation, its ability to learn dexterous skills on novel embodiments with limited data, and its capacity to complete complex, long-horizon tasks using only direct language prompting, such as laundry folding. In all settings, our method demonstrates superior performance compared to state-of-the-art models like Octo, OpenVLA, and Diffusion Policy.

Zhongyi Zhou, Yichen Zhu, Minjie Zhu et al.

Humans possess a unified cognitive ability to perceive, comprehend, and interact with the physical world. Why can't large language models replicate this holistic understanding? Through a systematic analysis of existing training paradigms in vision-language-action models (VLA), we identify two key challenges: spurious forgetting, where robot training overwrites crucial visual-text alignments, and task interference, where competing control and understanding tasks degrade performance when trained jointly. To overcome these limitations, we propose ChatVLA, a novel framework featuring Phased Alignment Training, which incrementally integrates multimodal data after initial control mastery, and a Mixture-of-Experts architecture to minimize task interference. ChatVLA demonstrates competitive performance on visual question-answering datasets and significantly surpasses state-of-the-art vision-language-action (VLA) methods on multimodal understanding benchmarks. Notably, it achieves a six times higher performance on MMMU and scores 47.2% on MMStar with a more parameter-efficient design than ECoT. Furthermore, ChatVLA demonstrates superior performance on 25 real-world robot manipulation tasks compared to existing VLA methods like OpenVLA. Our findings highlight the potential of our unified framework for achieving both robust multimodal understanding and effective robot control.

Chengmeng Li, Junjie Wen, Yan Peng et al.

Vision-Language-Action (VLA) models excel at robotic tasks by leveraging large-scale 2D vision-language pretraining, but their reliance on RGB images limits spatial reasoning critical for real-world interaction. Retraining these models with 3D data is computationally prohibitive, while discarding existing 2D datasets wastes valuable resources. To bridge this gap, we propose PointVLA, a framework that enhances pre-trained VLAs with point cloud inputs without requiring retraining. Our method freezes the vanilla action expert and injects 3D features via a lightweight modular block. To identify the most effective way of integrating point cloud representations, we conduct a skip-block analysis to pinpoint less useful blocks in the vanilla action expert, ensuring that 3D features are injected only into these blocks--minimizing disruption to pre-trained representations. Extensive experiments demonstrate that PointVLA outperforms state-of-the-art 2D imitation learning methods, such as OpenVLA, Diffusion Policy and DexVLA, across both simulated and real-world robotic tasks. Specifically, we highlight several key advantages of PointVLA enabled by point cloud integration: (1) Few-shot multi-tasking, where PointVLA successfully performs four different tasks using only 20 demonstrations each; (2) Real-vs-photo discrimination, where PointVLA distinguishes real objects from their images, leveraging 3D world knowledge to improve safety and reliability; (3) Height adaptability, Unlike conventional 2D imitation learning methods, PointVLA enables robots to adapt to objects at varying table height that unseen in train data. Furthermore, PointVLA achieves strong performance in long-horizon tasks, such as picking and packing objects from a moving conveyor belt, showcasing its ability to generalize across complex, dynamic environments.

Junjie Wen, Minjie Zhu, Yichen Zhu et al.

In this paper, we present DiffusionVLA, a novel framework that seamlessly combines the autoregression model with the diffusion model for learning visuomotor policy. Central to our approach is a next-token prediction objective, enabling the model to reason effectively over the user's query in the context of current observations. Subsequently, a diffusion model is attached to generate robust action outputs. To enhance policy learning through self-reasoning, we introduce a novel reasoning injection module that integrates reasoning phrases directly into the policy learning process. The whole framework is simple and flexible, making it easy to deploy and upgrade. We conduct extensive experiments using multiple real robots to validate the effectiveness of DiffusionVLA. Our tests include a challenging factory sorting task, where DiffusionVLA successfully categorizes objects, including those not seen during training. We observe that the reasoning module makes the model interpretable. It allows observers to understand the model thought process and identify potential causes of policy failures. Additionally, we test DiffusionVLA on a zero-shot bin-picking task, achieving 63.7\% accuracy on 102 previously unseen objects. Our method demonstrates robustness to visual changes, such as distractors and new backgrounds, and easily adapts to new embodiments. Furthermore, DiffusionVLA can follow novel instructions and retain conversational ability. Notably, DiffusionVLA is data-efficient and fast at inference; our smallest DiffusionVLA-2B runs 82Hz on a single A6000 GPU and can train from scratch on less than 50 demonstrations for a complex task. Finally, we scale the model from 2B to 72B parameters, showcasing improved generalization capabilities with increased model size.

Minjie Zhu, Yichen Zhu, Jinming Li et al.

The language-conditioned robotic manipulation aims to transfer natural language instructions into executable actions, from simple pick-and-place to tasks requiring intent recognition and visual reasoning. Inspired by the dual process theory in cognitive science, which suggests two parallel systems of fast and slow thinking in human decision-making, we introduce Robotics with Fast and Slow Thinking (RFST), a framework that mimics human cognitive architecture to classify tasks and makes decisions on two systems based on instruction types. Our RFST consists of two key components: 1) an instruction discriminator to determine which system should be activated based on the current user instruction, and 2) a slow-thinking system that is comprised of a fine-tuned vision language model aligned with the policy networks, which allows the robot to recognize user intention or perform reasoning tasks. To assess our methodology, we built a dataset featuring real-world trajectories, capturing actions ranging from spontaneous impulses to tasks requiring deliberate contemplation. Our results, both in simulation and real-world scenarios, confirm that our approach adeptly manages intricate tasks that demand intent recognition and reasoning. The project is available at https://jlm-z.github.io/RSFT/

Junjie Wen, Yichen Zhu, Minjie Zhu et al.

Humans interpret scenes by recognizing both the identities and positions of objects in their observations. For a robot to perform tasks such as \enquote{pick and place}, understanding both what the objects are and where they are located is crucial. While the former has been extensively discussed in the literature that uses the large language model to enrich the text descriptions, the latter remains underexplored. In this work, we introduce the \textit{Object-Centric Instruction Augmentation (OCI)} framework to augment highly semantic and information-dense language instruction with position cues. We utilize a Multi-modal Large Language Model (MLLM) to weave knowledge of object locations into natural language instruction, thus aiding the policy network in mastering actions for versatile manipulation. Additionally, we present a feature reuse mechanism to integrate the vision-language features from off-the-shelf pre-trained MLLM into policy networks. Through a series of simulated and real-world robotic tasks, we demonstrate that robotic manipulator imitation policies trained with our enhanced instructions outperform those relying solely on traditional language instructions.

Minjie Zhu, Yichen Zhu, Jinming Li et al.

Imitation learning has proven to be highly effective in teaching robots dexterous manipulation skills. However, it typically relies on large amounts of human demonstration data, which limits its scalability and applicability in dynamic, real-world environments. One key challenge in this context is object generalization, where a robot trained to perform a task with one object, such as "hand over the apple," struggles to transfer its skills to a semantically similar but visually different object, such as "hand over the peach." This gap in generalization to new objects beyond those in the same category has yet to be adequately addressed in previous work on end-to-end visuomotor policy learning. In this paper, we present a simple yet effective approach for achieving object generalization through Vision-Language-Action (VLA) models, referred to as \textbf{ObjectVLA}. Our model enables robots to generalize learned skills to novel objects without requiring explicit human demonstrations for each new target object. By leveraging vision-language pair data, our method provides a lightweight and scalable way to inject knowledge about the target object, establishing an implicit link between the object and the desired action. We evaluate ObjectVLA on a real robotic platform, demonstrating its ability to generalize across 100 novel objects with a 64\% success rate in selecting objects not seen during training. Furthermore, we propose a more accessible method for enhancing object generalization in VLA models, using a smartphone to capture a few images and fine-tune the pre-trained model. These results highlight the effectiveness of our approach in enabling object-level generalization and reducing the need for extensive human demonstrations, paving the way for more flexible and scalable robotic learning systems.

Yuqian Li, Yupei Du, Yufang Liu et al.

Language models excel in various tasks by making complex decisions, yet understanding the rationale behind these decisions remains a challenge. This paper investigates \emph{data-centric interpretability} in language models, focusing on the next-word prediction task. Using representer theorem, we identify two types of \emph{support samples}-those that either promote or deter specific predictions. Our findings reveal that being a support sample is an intrinsic property, predictable even before training begins. Additionally, while non-support samples are less influential in direct predictions, they play a critical role in preventing overfitting and shaping generalization and representation learning. Notably, the importance of non-support samples increases in deeper layers, suggesting their significant role in intermediate representation formation. These insights shed light on the interplay between data and model decisions, offering a new dimension to understanding language model behavior and interpretability.

Yichen Zhu, Yuqin Zhu, Jie Du et al.

The vision transformer splits each image into a sequence of tokens with fixed length and processes the tokens in the same way as words in natural language processing. More tokens normally lead to better performance but considerably increased computational cost. Motivated by the proverb "A picture is worth a thousand words" we aim to accelerate the ViT model by making a long image short. To this end, we propose a novel approach to assign token length adaptively during inference. Specifically, we first train a ViT model, called Resizable-ViT (ReViT), that can process any given input with diverse token lengths. Then, we retrieve the "token-length label" from ReViT and use it to train a lightweight Token-Length Assigner (TLA). The token-length labels are the smallest number of tokens to split an image that the ReViT can make the correct prediction, and TLA is learned to allocate the optimal token length based on these labels. The TLA enables the ReViT to process the image with the minimum sufficient number of tokens during inference. Thus, the inference speed is boosted by reducing the token numbers in the ViT model. Our approach is general and compatible with modern vision transformer architectures and can significantly reduce computational expanse. We verified the effectiveness of our methods on multiple representative ViT models (DeiT, LV-ViT, and TimesFormer) across two tasks (image classification and action recognition).

Yichen Zhu, Jie Du, Yuqin Zhu et al.

This work investigates the usage of batch normalization in neural architecture search (NAS). Specifically, Frankle et al. find that training BatchNorm only can achieve nontrivial performance. Furthermore, Chen et al. claim that training BatchNorm only can speed up the training of the one-shot NAS supernet over ten times. Critically, there is no effort to understand 1) why training BatchNorm only can find the perform-well architectures with the reduced supernet-training time, and 2) what is the difference between the train-BN-only supernet and the standard-train supernet. We begin by showing that the train-BN-only networks converge to the neural tangent kernel regime, obtain the same training dynamics as train all parameters theoretically. Our proof supports the claim to train BatchNorm only on supernet with less training time. Then, we empirically disclose that train-BN-only supernet provides an advantage on convolutions over other operators, cause unfair competition between architectures. This is due to only the convolution operator being attached with BatchNorm. Through experiments, we show that such unfairness makes the search algorithm prone to select models with convolutions. To solve this issue, we introduce fairness in the search space by placing a BatchNorm layer on every operator. However, we observe that the performance predictor in Chen et al. is inapplicable on the new search space. To this end, we propose a novel composite performance indicator to evaluate networks from three perspectives: expressivity, trainability, and uncertainty, derived from the theoretical property of BatchNorm. We demonstrate the effectiveness of our approach on multiple NAS-benchmarks (NAS-Bench101, NAS-Bench-201) and search spaces (DARTS search space and MobileNet search space).