Shubham Agrawal

Isaac Kasahara, Shubham Agrawal, Selim Engin et al. · apple-ml, cmu

General scene reconstruction refers to the task of estimating the full 3D geometry and texture of a scene containing previously unseen objects. In many practical applications such as AR/VR, autonomous navigation, and robotics, only a single view of the scene may be available, making the scene reconstruction task challenging. In this paper, we present a method for scene reconstruction by structurally breaking the problem into two steps: rendering novel views via inpainting and 2D to 3D scene lifting. Specifically, we leverage the generalization capability of large visual language models (Dalle-2) to inpaint the missing areas of scene color images rendered from different views. Next, we lift these inpainted images to 3D by predicting normals of the inpainted image and solving for the missing depth values. By predicting for normals instead of depth directly, our method allows for robustness to changes in depth distributions and scale. With rigorous quantitative evaluation, we show that our method outperforms multiple baselines while providing generalization to novel objects and scenes.

Mert Cemri, Shubham Agrawal, Akshat Gupta et al.

The paradigm of automated program generation is shifting from one-shot generation to inference-time search, where Large Language Models (LLMs) function as semantic mutation operators within evolutionary loops. While effective, these systems are currently governed by static schedules that fail to account for the non-stationary dynamics of the search process. This rigidity results in substantial computational waste, as resources are indiscriminately allocated to stagnating populations while promising frontiers remain under-exploited. We introduce AdaEvolve, a framework that reformulates LLM-driven evolution as a hierarchical adaptive optimization problem. AdaEvolve uses an "accumulated improvement signal" to unify decisions across three levels: Local Adaptation, which dynamically modulates the exploration intensity within a population of solution candidates; Global Adaptation, which routes the global resource budget via bandit-based scheduling across different solution candidate populations; and Meta-Guidance which generates novel solution tactics based on the previously generated solutions and their corresponding improvements when the progress stalls. We demonstrate that AdaEvolve consistently outperforms the open-sourced baselines across 185 different open-ended optimization problems including combinatorial, systems optimization and algorithm design problems.

Yong Zhao, Runxin He, Nicholas Kersting et al.

Understanding why a neural network model makes certain decisions can be as important as the inference performance. Various methods have been proposed to help practitioners explain the prediction of a neural network model, of which Shapley values are most popular. SHAP package is a leading implementation of Shapley values to explain neural networks implemented in TensorFlow or PyTorch but lacks cross-platform support, one-shot deployment and is highly inefficient. To address these problems, we present the ONNXExplainer, which is a generic framework to explain neural networks using Shapley values in the ONNX ecosystem. In ONNXExplainer, we develop its own automatic differentiation and optimization approach, which not only enables One-Shot Deployment of neural networks inference and explanations, but also significantly improves the efficiency to compute explanation with less memory consumption. For fair comparison purposes, we also implement the same optimization in TensorFlow and PyTorch and measure its performance against the current state of the art open-source counterpart, SHAP. Extensive benchmarks demonstrate that the proposed optimization approach improves the explanation latency of VGG19, ResNet50, DenseNet201, and EfficientNetB0 by as much as 500%.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

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.

Mansi Goel, Ayush Agarwal, Shubham Agrawal et al.

Food touches our lives through various endeavors, including flavor, nourishment, health, and sustainability. Recipes are cultural capsules transmitted across generations via unstructured text. Automated protocols for recognizing named entities, the building blocks of recipe text, are of immense value for various applications ranging from information extraction to novel recipe generation. Named entity recognition is a technique for extracting information from unstructured or semi-structured data with known labels. Starting with manually-annotated data of 6,611 ingredient phrases, we created an augmented dataset of 26,445 phrases cumulatively. Simultaneously, we systematically cleaned and analyzed ingredient phrases from RecipeDB, the gold-standard recipe data repository, and annotated them using the Stanford NER. Based on the analysis, we sampled a subset of 88,526 phrases using a clustering-based approach while preserving the diversity to create the machine-annotated dataset. A thorough investigation of NER approaches on these three datasets involving statistical, fine-tuning of deep learning-based language models and few-shot prompting on large language models (LLMs) provides deep insights. We conclude that few-shot prompting on LLMs has abysmal performance, whereas the fine-tuned spaCy-transformer emerges as the best model with macro-F1 scores of 95.9%, 96.04%, and 95.71% for the manually-annotated, augmented, and machine-annotated datasets, respectively.

Shubham Agrawal, Prasang Gupta

The rapid growth of large language models (LLMs) with diverse capabilities, latency and computational costs presents a critical deployment challenge: selecting the most suitable model for each prompt to optimize the trade-off between performance and efficiency. We introduce LLMRank, a prompt-aware routing framework that leverages rich, human-readable features extracted from prompts, including task type, reasoning patterns, complexity indicators, syntactic cues, and signals from a lightweight proxy solver. Unlike prior one-shot routers that rely solely on latent embeddings, LLMRank predicts per-model utility using a neural ranking model trained on RouterBench, comprising 36,497 prompts spanning 11 benchmarks and 11 state-of-the-art LLMs, from small efficient models to large frontier systems. Our approach achieves up to 89.2% of oracle utility, while providing interpretable feature attributions that explain routing decisions. Extensive studies demonstrate the importance of multifaceted feature extraction and the hybrid ranking objective, highlighting the potential of feature-driven routing for efficient and transparent LLM deployment.

CodeGemma Team, Heri Zhao, Jeffrey Hui et al.

This paper introduces CodeGemma, a collection of specialized open code models built on top of Gemma, capable of a variety of code and natural language generation tasks. We release three model variants. CodeGemma 7B pretrained (PT) and instruction-tuned (IT) variants have remarkably resilient natural language understanding, excel in mathematical reasoning, and match code capabilities of other open models. CodeGemma 2B is a state-of-the-art code completion model designed for fast code infilling and open-ended generation in latency-sensitive settings.

Sergei Krutikov, Bulat Khaertdinov, Rodion Kiriukhin et al.

Transformer-based neural networks, empowered by Self-Supervised Learning (SSL), have demonstrated unprecedented performance across various domains. However, related literature suggests that tabular Transformers may struggle to outperform classical Machine Learning algorithms, such as Gradient Boosted Decision Trees (GBDT). In this paper, we aim to challenge GBDTs with tabular Transformers on a typical task faced in e-commerce, namely fraud detection. Our study is additionally motivated by the problem of selection bias, often occurring in real-life fraud detection systems. It is caused by the production system affecting which subset of traffic becomes labeled. This issue is typically addressed by sampling randomly a small part of the whole production data, referred to as a Control Group. This subset follows a target distribution of production data and therefore is usually preferred for training classification models with standard ML algorithms. Our methodology leverages the capabilities of Transformers to learn transferable representations using all available data by means of SSL, giving it an advantage over classical methods. Furthermore, we conduct large-scale experiments, pre-training tabular Transformers on vast amounts of data instances and fine-tuning them on smaller target datasets. The proposed approach outperforms heavily tuned GBDTs by a considerable margin of the Average Precision (AP) score in offline evaluations. Finally, we report the results of an online A/B experiment. Experimental results confirm the superiority of tabular Transformers compared to GBDTs in production, demonstrated by a statistically significant improvement in our business metric.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Shubham Agrawal, Nikhil Chavan-Dafle, Isaac Kasahara et al.

Robotic manipulation systems operating in complex environments rely on perception systems that provide information about the geometry (pose and 3D shape) of the objects in the scene along with other semantic information such as object labels. This information is then used for choosing the feasible grasps on relevant objects. In this paper, we present a novel method to provide this geometric and semantic information of all objects in the scene as well as feasible grasps on those objects simultaneously. The main advantage of our method is its speed as it avoids sequential perception and grasp planning steps. With detailed quantitative analysis, we show that our method delivers competitive performance compared to the state-of-the-art dedicated methods for object shape, pose, and grasp predictions while providing fast inference at 30 frames per second speed.

Georg Hille, Shubham Agrawal, Pavan Tummala et al.

Deep learning-based segmentation of the liver and hepatic lesions therein steadily gains relevance in clinical practice due to the increasing incidence of liver cancer each year. Whereas various network variants with overall promising results in the field of medical image segmentation have been successfully developed over the last years, almost all of them struggle with the challenge of accurately segmenting hepatic lesions in magnetic resonance imaging (MRI). This led to the idea of combining elements of convolutional and transformer-based architectures to overcome the existing limitations. This work presents a hybrid network called SWTR-Unet, consisting of a pretrained ResNet, transformer blocks as well as a common Unet-style decoder path. This network was primarily applied to single-modality non-contrast-enhanced liver MRI and additionally to the publicly available computed tomography (CT) data of the liver tumor segmentation (LiTS) challenge to verify the applicability on other modalities. For a broader evaluation, multiple state-of-the-art networks were implemented and applied, ensuring a direct comparability. Furthermore, correlation analysis and an ablation study were carried out, to investigate various influencing factors on the segmentation accuracy of the presented method. With Dice scores of averaged 98+-2% for liver and 81+-28% lesion segmentation on the MRI dataset and 97+-2% and 79+-25%, respectively on the CT dataset, the proposed SWTR-Unet proved to be a precise approach for liver and hepatic lesion segmentation with state-of-the-art results for MRI and competing accuracy in CT imaging. The achieved segmentation accuracy was found to be on par with manually performed expert segmentations as indicated by inter-observer variabilities for liver lesion segmentation. In conclusion, the presented method could save valuable time and resources in clinical practice.

Yulong Li, Shubham Agrawal, Jen-Shuo Liu et al.

Studies in robot teleoperation have been centered around action specifications -- from continuous joint control to discrete end-effector pose control. However, these robot-centric interfaces often require skilled operators with extensive robotics expertise. To make teleoperation accessible to non-expert users, we propose the framework "Scene Editing as Teleoperation" (SEaT), where the key idea is to transform the traditional "robot-centric" interface into a "scene-centric" interface -- instead of controlling the robot, users focus on specifying the task's goal by manipulating digital twins of the real-world objects. As a result, a user can perform teleoperation without any expert knowledge of the robot hardware. To achieve this goal, we utilize a category-agnostic scene-completion algorithm that translates the real-world workspace (with unknown objects) into a manipulable virtual scene representation and an action-snapping algorithm that refines the user input before generating the robot's action plan. To train the algorithms, we procedurally generated a large-scale, diverse kit-assembly dataset that contains object-kit pairs that mimic real-world object-kitting tasks. Our experiments in simulation and on a real-world system demonstrate that our framework improves both the efficiency and success rate for 6DoF kit-assembly tasks. A user study demonstrates that SEaT framework participants achieve a higher task success rate and report a lower subjective workload compared to an alternative robot-centric interface. Video can be found at https://www.youtube.com/watch?v=-NdR3mkPbQQ .

Nikhil Chavan-Dafle, Sergiy Popovych, Shubham Agrawal et al.

Being able to grasp objects is a fundamental component of most robotic manipulation systems. In this paper, we present a new approach to simultaneously reconstruct a mesh and a dense grasp quality map of an object from a depth image. At the core of our approach is a novel camera-centric object representation called the "object shell" which is composed of an observed "entry image" and a predicted "exit image". We present an image-to-image residual ConvNet architecture in which the object shell and a grasp-quality map are predicted as separate output channels. The main advantage of the shell representation and the corresponding neural network architecture, ShellGrasp-Net, is that the input-output pixel correspondences in the shell representation are explicitly represented in the architecture. We show that this coupling yields superior generalization capabilities for object reconstruction and accurate grasp quality estimation implicitly considering the object geometry. Our approach yields an efficient dense grasp quality map and an object geometry estimate in a single forward pass. Both of these outputs can be used in a wide range of robotic manipulation applications. With rigorous experimental validation, both in simulation and on a real setup, we show that our shell-based method can be used to generate precise grasps and the associated grasp quality with over 90% accuracy. Diverse grasps computed on shell reconstructions allow the robot to select and execute grasps in cluttered scenes with more than 93% success rate.

Zhenjia Xu, Beichun Qi, Shubham Agrawal et al.

This paper aims to improve robots' versatility and adaptability by allowing them to use a large variety of end-effector tools and quickly adapt to new tools. We propose AdaGrasp, a method to learn a single grasping policy that generalizes to novel grippers. By training on a large collection of grippers, our algorithm is able to acquire generalizable knowledge of how different grippers should be used in various tasks. Given a visual observation of the scene and the gripper, AdaGrasp infers the possible grasp poses and their grasp scores by computing the cross convolution between the shape encodings of the gripper and scene. Intuitively, this cross convolution operation can be considered as an efficient way of exhaustively matching the scene geometry with gripper geometry under different grasp poses (i.e., translations and orientations), where a good "match" of 3D geometry will lead to a successful grasp. We validate our methods in both simulation and real-world environments. Our experiment shows that AdaGrasp significantly outperforms the existing multi-gripper grasping policy method, especially when handling cluttered environments and partial observations. Video is available at https://youtu.be/kknTYTbORfs

Huy Ha, Shubham Agrawal, Shuran Song

The 3D shape of a robot's end-effector plays a critical role in determining it's functionality and overall performance. Many industrial applications rely on task-specific gripper designs to ensure the system's robustness and accuracy. However, the process of manual hardware design is both costly and time-consuming, and the quality of the resulting design is dependent on the engineer's experience and domain expertise, which can easily be out-dated or inaccurate. The goal of this work is to use machine learning algorithms to automate the design of task-specific gripper fingers. We propose Fit2Form, a 3D generative design framework that generates pairs of finger shapes to maximize design objectives (i.e., grasp success, stability, and robustness) for target grasp objects. We model the design objectives by training a Fitness network to predict their values for pairs of gripper fingers and their corresponding grasp objects. This Fitness network then provides supervision to a 3D Generative network that produces a pair of 3D finger geometries for the target grasp object. Our experiments demonstrate that the proposed 3D generative design framework generates parallel jaw gripper finger shapes that achieve more stable and robust grasps compared to other general-purpose and task-specific gripper design algorithms. Video can be found at https://youtu.be/utKHP3qb1bg.

Shubham Agrawal, Anuj Pahuja, Simon Lucey

What's the most accurate 3D model of your face you can obtain while sitting at your desk? We attempt to answer this question in our work. High fidelity face reconstructions have so far been limited to either studio settings or through expensive 3D scanners. On the other hand, unconstrained reconstruction methods are typically limited by low-capacity models. Our method reconstructs accurate face geometry of a subject using a video shot from a smartphone in an unconstrained environment. Our approach takes advantage of recent advances in visual SLAM, keypoint detection, and object detection to improve accuracy and robustness. By not being constrained to a model subspace, our reconstructed meshes capture important details while being robust to noise and being topologically consistent. Our evaluations show that our method outperforms current single and multi-view baselines by a significant margin, both in terms of geometric accuracy and in capturing person-specific details important for making realistic looking models.

David Millard, Eric Heiden, Shubham Agrawal et al.

A key ingredient to achieving intelligent behavior is physical understanding that equips robots with the ability to reason about the effects of their actions in a dynamic environment. Several methods have been proposed to learn dynamics models from data that inform model-based control algorithms. While such learning-based approaches can model locally observed behaviors, they fail to generalize to more complex dynamics and under long time horizons. In this work, we introduce a differentiable physics simulator for rigid body dynamics. Leveraging various techniques for differential equation integration and gradient calculation, we compare different methods for parameter estimation that allow us to infer the simulation parameters that are relevant to estimation and control of physical systems. In the context of trajectory optimization, we introduce a closed-loop model-predictive control algorithm that infers the simulation parameters through experience while achieving cost-minimizing performance.

Gaurav Mittal, Shubham Agrawal, Anuva Agarwal et al.

Recent years have witnessed some exciting developments in the domain of generating images from scene-based text descriptions. These approaches have primarily focused on generating images from a static text description and are limited to generating images in a single pass. They are unable to generate an image interactively based on an incrementally additive text description (something that is more intuitive and similar to the way we describe an image). We propose a method to generate an image incrementally based on a sequence of graphs of scene descriptions (scene-graphs). We propose a recurrent network architecture that preserves the image content generated in previous steps and modifies the cumulative image as per the newly provided scene information. Our model utilizes Graph Convolutional Networks (GCN) to cater to variable-sized scene graphs along with Generative Adversarial image translation networks to generate realistic multi-object images without needing any intermediate supervision during training. We experiment with Coco-Stuff dataset which has multi-object images along with annotations describing the visual scene and show that our model significantly outperforms other approaches on the same dataset in generating visually consistent images for incrementally growing scene graphs.

Tejas Khot, Shubham Agrawal, Shubham Tulsiani et al.

We present a learning based approach for multi-view stereopsis (MVS). While current deep MVS methods achieve impressive results, they crucially rely on ground-truth 3D training data, and acquisition of such precise 3D geometry for supervision is a major hurdle. Our framework instead leverages photometric consistency between multiple views as supervisory signal for learning depth prediction in a wide baseline MVS setup. However, naively applying photo consistency constraints is undesirable due to occlusion and lighting changes across views. To overcome this, we propose a robust loss formulation that: a) enforces first order consistency and b) for each point, selectively enforces consistency with some views, thus implicitly handling occlusions. We demonstrate our ability to learn MVS without 3D supervision using a real dataset, and show that each component of our proposed robust loss results in a significant improvement. We qualitatively observe that our reconstructions are often more complete than the acquired ground truth, further showing the merits of this approach. Lastly, our learned model generalizes to novel settings, and our approach allows adaptation of existing CNNs to datasets without ground-truth 3D by unsupervised finetuning. Project webpage: https://tejaskhot.github.io/unsup_mvs

Swaminathan Gurumurthy, Shubham Agrawal

Semantic shape completion is a challenging problem in 3D computer vision where the task is to generate a complete 3D shape using a partial 3D shape as input. We propose a learning-based approach to complete incomplete 3D shapes through generative modeling and latent manifold optimization. Our algorithm works directly on point clouds. We use an autoencoder and a GAN to learn a distribution of embeddings for point clouds of object classes. An input point cloud with missing regions is first encoded to a feature vector. The representations learnt by the GAN are then used to find the best latent vector on the manifold using a combined optimization that finds a vector in the manifold of plausible vectors that is close to the original input (both in the feature space and the output space of the decoder). Experiments show that our algorithm is capable of successfully reconstructing point clouds with large missing regions with very high fidelity without having to rely on exemplar based database retrieval.