Wentao Yuan

Wentao Yuan, Adithyavairavan Murali, Arsalan Mousavian et al. · uw

With the advent of large language models and large-scale robotic datasets, there has been tremendous progress in high-level decision-making for object manipulation. These generic models are able to interpret complex tasks using language commands, but they often have difficulties generalizing to out-of-distribution objects due to the inability of low-level action primitives. In contrast, existing task-specific models excel in low-level manipulation of unknown objects, but only work for a single type of action. To bridge this gap, we present M2T2, a single model that supplies different types of low-level actions that work robustly on arbitrary objects in cluttered scenes. M2T2 is a transformer model which reasons about contact points and predicts valid gripper poses for different action modes given a raw point cloud of the scene. Trained on a large-scale synthetic dataset with 128K scenes, M2T2 achieves zero-shot sim2real transfer on the real robot, outperforming the baseline system with state-of-the-art task-specific models by about 19% in overall performance and 37.5% in challenging scenes where the object needs to be re-oriented for collision-free placement. M2T2 also achieves state-of-the-art results on a subset of language conditioned tasks in RLBench. Videos of robot experiments on unseen objects in both real world and simulation are available on our project website https://m2-t2.github.io.

Wentao Yuan, Qingtian Zhu, Xiangyue Liu et al.

Recently, Implicit Neural Representations (INRs) parameterized by neural networks have emerged as a powerful and promising tool to represent different kinds of signals due to its continuous, differentiable properties, showing superiorities to classical discretized representations. However, the training of neural networks for INRs only utilizes input-output pairs, and the derivatives of the target output with respect to the input, which can be accessed in some cases, are usually ignored. In this paper, we propose a training paradigm for INRs whose target output is image pixels, to encode image derivatives in addition to image values in the neural network. Specifically, we use finite differences to approximate image derivatives. We show how the training paradigm can be leveraged to solve typical INRs problems, i.e., image regression and inverse rendering, and demonstrate this training paradigm can improve the data-efficiency and generalization capabilities of INRs. The code of our method is available at \url{https://github.com/megvii-research/Sobolev_INRs}.

Yikang Ding, Qingtian Zhu, Xiangyue Liu et al.

Supervised multi-view stereo (MVS) methods have achieved remarkable progress in terms of reconstruction quality, but suffer from the challenge of collecting large-scale ground-truth depth. In this paper, we propose a novel self-supervised training pipeline for MVS based on knowledge distillation, termed KD-MVS, which mainly consists of self-supervised teacher training and distillation-based student training. Specifically, the teacher model is trained in a self-supervised fashion using both photometric and featuremetric consistency. Then we distill the knowledge of the teacher model to the student model through probabilistic knowledge transferring. With the supervision of validated knowledge, the student model is able to outperform its teacher by a large margin. Extensive experiments performed on multiple datasets show our method can even outperform supervised methods.

Chahyon Ku, Carl Winge, Ryan Diaz et al.

This paper primarily focuses on evaluating and benchmarking the robustness of visual representations in the context of object assembly tasks. Specifically, it investigates the alignment and insertion of objects with geometrical extrusions and intrusions, commonly referred to as a peg-in-hole task. The accuracy required to detect and orient the peg and the hole geometry in SE(3) space for successful assembly poses significant challenges. Addressing this, we employ a general framework in visuomotor policy learning that utilizes visual pretraining models as vision encoders. Our study investigates the robustness of this framework when applied to a dual-arm manipulation setup, specifically to the grasp variations. Our quantitative analysis shows that existing pretrained models fail to capture the essential visual features necessary for this task. However, a visual encoder trained from scratch consistently outperforms the frozen pretrained models. Moreover, we discuss rotation representations and associated loss functions that substantially improve policy learning. We present a novel task scenario designed to evaluate the progress in visuomotor policy learning, with a specific focus on improving the robustness of intricate assembly tasks that require both geometrical and spatial reasoning. Videos, additional experiments, dataset, and code are available at https://bit.ly/geometric-peg-in-hole .

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Long Cheng, Jiafei Duan, Yi Ru Wang et al. · uw

Pointing serves as a fundamental and intuitive mechanism for grounding language within visual contexts, with applications spanning robotics, assistive technologies, and interactive AI systems. While recent multimodal models have started to support pointing capabilities, existing benchmarks typically focus only on referential object localization tasks. We introduce PointArena, a comprehensive platform for evaluating multimodal pointing across diverse reasoning scenarios. PointArena comprises three components: (1) Point-Bench, a curated dataset containing approximately 1,000 pointing tasks across five reasoning categories; (2) Point-Battle, an interactive, web-based arena facilitating blind, pairwise model comparisons, which has already gathered over 4,500 anonymized votes; and (3) Point-Act, a real-world robotic manipulation system allowing users to directly evaluate multimodal model pointing capabilities in practical settings. We conducted extensive evaluations of both state-of-the-art open-source and proprietary multimodal models. Results indicate that Molmo-72B consistently outperforms other models, though proprietary models increasingly demonstrate comparable performance. Additionally, we find that supervised training specifically targeting pointing tasks significantly enhances model performance. Across our multi-stage evaluation pipeline, we also observe strong correlations, underscoring the critical role of precise pointing capabilities in enabling multimodal models to effectively bridge abstract reasoning with concrete, real-world actions. Project page: https://pointarena.github.io/

Yikang Ding, Wentao Yuan, Qingtian Zhu et al.

In this paper, we present TransMVSNet, based on our exploration of feature matching in multi-view stereo (MVS). We analogize MVS back to its nature of a feature matching task and therefore propose a powerful Feature Matching Transformer (FMT) to leverage intra- (self-) and inter- (cross-) attention to aggregate long-range context information within and across images. To facilitate a better adaptation of the FMT, we leverage an Adaptive Receptive Field (ARF) module to ensure a smooth transit in scopes of features and bridge different stages with a feature pathway to pass transformed features and gradients across different scales. In addition, we apply pair-wise feature correlation to measure similarity between features, and adopt ambiguity-reducing focal loss to strengthen the supervision. To the best of our knowledge, TransMVSNet is the first attempt to leverage Transformer into the task of MVS. As a result, our method achieves state-of-the-art performance on DTU dataset, Tanks and Temples benchmark, and BlendedMVS dataset. The code of our method will be made available at https://github.com/MegviiRobot/TransMVSNet .

Adithyavairavan Murali, Balakumar Sundaralingam, Yu-Wei Chao et al. · nvidia, uw

Grasping is a fundamental robot skill, yet despite significant research advancements, learning-based 6-DOF grasping approaches are still not turnkey and struggle to generalize across different embodiments and in-the-wild settings. We build upon the recent success on modeling the object-centric grasp generation process as an iterative diffusion process. Our proposed framework, GraspGen, consists of a DiffusionTransformer architecture that enhances grasp generation, paired with an efficient discriminator to score and filter sampled grasps. We introduce a novel and performant on-generator training recipe for the discriminator. To scale GraspGen to both objects and grippers, we release a new simulated dataset consisting of over 53 million grasps. We demonstrate that GraspGen outperforms prior methods in simulations with singulated objects across different grippers, achieves state-of-the-art performance on the FetchBench grasping benchmark, and performs well on a real robot with noisy visual observations.

Chaoyuan Zhang, Zhaowei Li, Wentao Yuan

Integrating large language models (LLMs) into closed-loop robotic task planning has become increasingly popular within embodied artificial intelligence. Previous efforts mainly focused on leveraging the strong reasoning abilities of LLMs to enhance task planning performance while often overlooking task planning efficiency and executability due to repetitive queries to LLMs. This paper addresses the synergy between LLMs and task planning systems, aiming to minimize redundancy while enhancing planning effectiveness. Specifically, building upon Prog-Prompt and the high-level concept of Tree-Planner, we propose Vote-Tree-Planner. This sampling strategy utilizes votes to guide plan traversal during the decision-making process. Our approach is motivated by a straightforward observation: assigning weights to agents during decision-making enables the evaluation of critical paths before execution. With this simple vote-tree construction, our method further improves the success rate and reduces the number of queries to LLMs. The experimental results highlight that our Vote-Tree-Planner demonstrates greater stability and shows a higher average success rate and goal condition recall on the unseen dataset compared with previous baseline methods. These findings underscore the potential of the Vote-Tree-Planner to enhance planning accuracy, reliability, and efficiency in LLM-based planning systems.

Jiafei Duan, Wentao Yuan, Wilbert Pumacay et al.

Large-scale endeavors like and widespread community efforts such as Open-X-Embodiment have contributed to growing the scale of robot demonstration data. However, there is still an opportunity to improve the quality, quantity, and diversity of robot demonstration data. Although vision-language models have been shown to automatically generate demonstration data, their utility has been limited to environments with privileged state information, they require hand-designed skills, and are limited to interactions with few object instances. We propose Manipulate-Anything, a scalable automated generation method for real-world robotic manipulation. Unlike prior work, our method can operate in real-world environments without any privileged state information, hand-designed skills, and can manipulate any static object. We evaluate our method using two setups. First, Manipulate-Anything successfully generates trajectories for all 7 real-world and 14 simulation tasks, significantly outperforming existing methods like VoxPoser. Second, Manipulate-Anything's demonstrations can train more robust behavior cloning policies than training with human demonstrations, or from data generated by VoxPoser, Scaling-up, and Code-As-Policies. We believe Manipulate-Anything can be a scalable method for both generating data for robotics and solving novel tasks in a zero-shot setting. Project page: https://robot-ma.github.io/.

Wentao Yuan, Jiafei Duan, Valts Blukis et al.

From rearranging objects on a table to putting groceries into shelves, robots must plan precise action points to perform tasks accurately and reliably. In spite of the recent adoption of vision language models (VLMs) to control robot behavior, VLMs struggle to precisely articulate robot actions using language. We introduce an automatic synthetic data generation pipeline that instruction-tunes VLMs to robotic domains and needs. Using the pipeline, we train RoboPoint, a VLM that predicts image keypoint affordances given language instructions. Compared to alternative approaches, our method requires no real-world data collection or human demonstration, making it much more scalable to diverse environments and viewpoints. In addition, RoboPoint is a general model that enables several downstream applications such as robot navigation, manipulation, and augmented reality (AR) assistance. Our experiments demonstrate that RoboPoint outperforms state-of-the-art VLMs (GPT-4o) and visual prompting techniques (PIVOT) by 21.8% in the accuracy of predicting spatial affordance and by 30.5% in the success rate of downstream tasks. Project website: https://robo-point.github.io.

Wentao Yuan, Chris Paxton, Karthik Desingh et al.

Sequential manipulation tasks require a robot to perceive the state of an environment and plan a sequence of actions leading to a desired goal state. In such tasks, the ability to reason about spatial relations among object entities from raw sensor inputs is crucial in order to determine when a task has been completed and which actions can be executed. In this work, we propose SORNet (Spatial Object-Centric Representation Network), a framework for learning object-centric representations from RGB images conditioned on a set of object queries, represented as image patches called canonical object views. With only a single canonical view per object and no annotation, SORNet generalizes zero-shot to object entities whose shape and texture are both unseen during training. We evaluate SORNet on various spatial reasoning tasks such as spatial relation classification and relative direction regression in complex tabletop manipulation scenarios and show that SORNet significantly outperforms baselines including state-of-the-art representation learning techniques. We also demonstrate the application of the representation learned by SORNet on visual-servoing and task planning for sequential manipulation on a real robot.

Wentao Yuan, Zhaoyang Lv, Tanner Schmidt et al.

We present STaR, a novel method that performs Self-supervised Tracking and Reconstruction of dynamic scenes with rigid motion from multi-view RGB videos without any manual annotation. Recent work has shown that neural networks are surprisingly effective at the task of compressing many views of a scene into a learned function which maps from a viewing ray to an observed radiance value via volume rendering. Unfortunately, these methods lose all their predictive power once any object in the scene has moved. In this work, we explicitly model rigid motion of objects in the context of neural representations of radiance fields. We show that without any additional human specified supervision, we can reconstruct a dynamic scene with a single rigid object in motion by simultaneously decomposing it into its two constituent parts and encoding each with its own neural representation. We achieve this by jointly optimizing the parameters of two neural radiance fields and a set of rigid poses which align the two fields at each frame. On both synthetic and real world datasets, we demonstrate that our method can render photorealistic novel views, where novelty is measured on both spatial and temporal axes. Our factored representation furthermore enables animation of unseen object motion.

Wentao Yuan, Ben Eckart, Kihwan Kim et al.

Point cloud registration is a fundamental problem in 3D computer vision, graphics and robotics. For the last few decades, existing registration algorithms have struggled in situations with large transformations, noise, and time constraints. In this paper, we introduce Deep Gaussian Mixture Registration (DeepGMR), the first learning-based registration method that explicitly leverages a probabilistic registration paradigm by formulating registration as the minimization of KL-divergence between two probability distributions modeled as mixtures of Gaussians. We design a neural network that extracts pose-invariant correspondences between raw point clouds and Gaussian Mixture Model (GMM) parameters and two differentiable compute blocks that recover the optimal transformation from matched GMM parameters. This construction allows the network learn an SE(3)-invariant feature space, producing a global registration method that is real-time, generalizable, and robust to noise. Across synthetic and real-world data, our proposed method shows favorable performance when compared with state-of-the-art geometry-based and learning-based registration methods.

Wentao Yuan, David Held, Christoph Mertz et al.

3D point cloud is an efficient and flexible representation of 3D structures. Recently, neural networks operating on point clouds have shown superior performance on 3D understanding tasks such as shape classification and part segmentation. However, performance on such tasks is evaluated on complete shapes aligned in a canonical frame, while real world 3D data are partial and unaligned. A key challenge in learning from partial, unaligned point cloud data is to learn features that are invariant or equivariant with respect to geometric transformations. To address this challenge, we propose the Iterative Transformer Network (IT-Net), a network module that canonicalizes the pose of a partial object with a series of 3D rigid transformations predicted in an iterative fashion. We demonstrate the efficacy of IT-Net as an anytime pose estimator from partial point clouds without using complete object models. Further, we show that IT-Net achieves superior performance over alternative 3D transformer networks on various tasks, such as partial shape classification and object part segmentation.

Wentao Yuan, Tejas Khot, David Held et al.

Shape completion, the problem of estimating the complete geometry of objects from partial observations, lies at the core of many vision and robotics applications. In this work, we propose Point Completion Network (PCN), a novel learning-based approach for shape completion. Unlike existing shape completion methods, PCN directly operates on raw point clouds without any structural assumption (e.g. symmetry) or annotation (e.g. semantic class) about the underlying shape. It features a decoder design that enables the generation of fine-grained completions while maintaining a small number of parameters. Our experiments show that PCN produces dense, complete point clouds with realistic structures in the missing regions on inputs with various levels of incompleteness and noise, including cars from LiDAR scans in the KITTI dataset.