Hongbin Liu

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Xingjian Luo, You Pang, Zhen Chen et al.

Surgical phase recognition is crucial to providing surgery understanding in smart operating rooms. Despite great progress in automatic surgical phase recognition, most existing methods are still restricted by two problems. First, these methods cannot capture discriminative visual features for each frame and motion information with simple 2D networks. Second, the frame-by-frame recognition paradigm degrades the performance due to unstable predictions within each phase, termed as phase shaking. To address these two challenges, we propose a Surgical Phase LocAlization Network, named SurgPLAN, to facilitate a more accurate and stable surgical phase recognition with the principle of temporal detection. Specifically, we first devise a Pyramid SlowFast (PSF) architecture to serve as the visual backbone to capture multi-scale spatial and temporal features by two branches with different frame sampling rates. Moreover, we propose a Temporal Phase Localization (TPL) module to generate the phase prediction based on temporal region proposals, which ensures accurate and consistent predictions within each surgical phase. Extensive experiments confirm the significant advantages of our SurgPLAN over frame-by-frame approaches in terms of both accuracy and stability.

Zhen Sun, Huan Xu, Jinlin Wu et al.

In surgical procedures, correct instrument counting is essential. Instance segmentation is a location method that locates not only an object's bounding box but also each pixel's specific details. However, obtaining mask-level annotations is labor-intensive in instance segmentation. To address this issue, we propose a novel yet effective weakly-supervised surgical instrument instance segmentation approach, named Point-based Weakly-supervised Instance Segmentation (PWISeg). PWISeg adopts an FCN-based architecture with point-to-box and point-to-mask branches to model the relationships between feature points and bounding boxes, as well as feature points and segmentation masks on FPN, accomplishing instrument detection and segmentation jointly in a single model. Since mask level annotations are hard to available in the real world, for point-to-mask training, we introduce an unsupervised projection loss, utilizing the projected relation between predicted masks and bboxes as supervision signal. On the other hand, we annotate a few pixels as the key pixel for each instrument. Based on this, we further propose a key pixel association loss and a key pixel distribution loss, driving the point-to-mask branch to generate more accurate segmentation predictions. To comprehensively evaluate this task, we unveil a novel surgical instrument dataset with manual annotations, setting up a benchmark for further research. Our comprehensive research trial validated the superior performance of our PWISeg. The results show that the accuracy of surgical instrument segmentation is improved, surpassing most methods of instance segmentation via weakly supervised bounding boxes. This improvement is consistently observed in our proposed dataset and when applied to the public HOSPI-Tools dataset.

Yiming Huang, Beilei Cui, Long Bai et al.

Simultaneous Localization and Mapping (SLAM) is essential for precise surgical interventions and robotic tasks in minimally invasive procedures. While recent advancements in 3D Gaussian Splatting (3DGS) have improved SLAM with high-quality novel view synthesis and fast rendering, these systems struggle with accurate depth and surface reconstruction due to multi-view inconsistencies. Simply incorporating SLAM and 3DGS leads to mismatches between the reconstructed frames. In this work, we present Endo-2DTAM, a real-time endoscopic SLAM system with 2D Gaussian Splatting (2DGS) to address these challenges. Endo-2DTAM incorporates a surface normal-aware pipeline, which consists of tracking, mapping, and bundle adjustment modules for geometrically accurate reconstruction. Our robust tracking module combines point-to-point and point-to-plane distance metrics, while the mapping module utilizes normal consistency and depth distortion to enhance surface reconstruction quality. We also introduce a pose-consistent strategy for efficient and geometrically coherent keyframe sampling. Extensive experiments on public endoscopic datasets demonstrate that Endo-2DTAM achieves an RMSE of $1.87\pm 0.63$ mm for depth reconstruction of surgical scenes while maintaining computationally efficient tracking, high-quality visual appearance, and real-time rendering. Our code will be released at github.com/lastbasket/Endo-2DTAM.

Yiming Huang, Long Bai, Beilei Cui et al.

Accurate reconstruction of soft tissue is crucial for advancing automation in image-guided robotic surgery. The recent 3D Gaussian Splatting (3DGS) techniques and their variants, 4DGS, achieve high-quality renderings of dynamic surgical scenes in real-time. However, 3D-GS-based methods still struggle in scenarios with varying illumination, such as low light and over-exposure. Training 3D-GS in such extreme light conditions leads to severe optimization problems and devastating rendering quality. To address these challenges, we present Endo-4DGX, a novel reconstruction method with illumination-adaptive Gaussian Splatting designed specifically for endoscopic scenes with uneven lighting. By incorporating illumination embeddings, our method effectively models view-dependent brightness variations. We introduce a region-aware enhancement module to model the sub-area lightness at the Gaussian level and a spatial-aware adjustment module to learn the view-consistent brightness adjustment. With the illumination adaptive design, Endo-4DGX achieves superior rendering performance under both low-light and over-exposure conditions while maintaining geometric accuracy. Additionally, we employ an exposure control loss to restore the appearance from adverse exposure to the normal level for illumination-adaptive optimization. Experimental results demonstrate that Endo-4DGX significantly outperforms combinations of state-of-the-art reconstruction and restoration methods in challenging lighting environments, underscoring its potential to advance robot-assisted surgical applications. Our code is available at https://github.com/lastbasket/Endo-4DGX.

Song Lai, Haohan Zhao, Rong Feng et al.

Continual post-training (CPT) is a popular and effective technique for adapting foundation models like multimodal large language models to specific and ever-evolving downstream tasks. While existing research has primarily concentrated on methods like data replay, model expansion, or parameter regularization, the fundamental role of the learning paradigm within CPT remains largely unexplored. This paper presents a comparative analysis of two core post-training paradigms: supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), investigating their respective impacts on knowledge retention during CPT. Our experiments are conducted on a benchmark comprising seven diverse multimodal tasks, utilizing Qwen2.5-VL-7B-Instruct as the base model for continual post-training. The investigation yields two significant findings: (1) When continuously learning on downstream tasks, SFT leads to catastrophic forgetting of previously learned tasks. In contrast, RFT inherently preserves prior knowledge and achieve performance comparable to multi-task training. (2) RFT successfully protects and even enhances the model's general knowledge on standard benchmarks (e.g., MMMU and MMLU-Pro). Conversely, SFT degrades general model capabilities severely. Further analysis reveals that this stability is not primarily due to explicit mechanisms like KL penalty or chain-of-thought reasoning. Instead, we identify an implicit regularization mechanism inherent to RFT as a key contributing factor. Our theoretical analysis suggests that RFT's gradient updates are naturally scaled by the reward variance, acting as a data-dependent regularizer that inherently protects previously acquired knowledge. Finally, we propose a rollout-based instance filtering algorithm to enhance the stability and efficiency of RFT. Our comprehensive study demonstrates the superiority of RFT as a robust paradigm for continual post-training.

Huan Xu, Jinlin Wu, Guanglin Cao et al.

Ultrasound robots are increasingly used in medical diagnostics and early disease screening. However, current ultrasound robots lack the intelligence to understand human intentions and instructions, hindering autonomous ultrasound scanning. To solve this problem, we propose a novel Ultrasound Embodied Intelligence system that equips ultrasound robots with the large language model (LLM) and domain knowledge, thereby improving the efficiency of ultrasound robots. Specifically, we first design an ultrasound operation knowledge database to add expertise in ultrasound scanning to the LLM, enabling the LLM to perform precise motion planning. Furthermore, we devise a dynamic ultrasound scanning strategy based on a \textit{think-observe-execute} prompt engineering, allowing LLMs to dynamically adjust motion planning strategies during the scanning procedures. Extensive experiments demonstrate that our system significantly improves ultrasound scan efficiency and quality from verbal commands. This advancement in autonomous medical scanning technology contributes to non-invasive diagnostics and streamlined medical workflows.

Lihua Zhou, Mao Ye, Nianxin Li et al.

Deep learning often struggles when training and test data distributions differ. Traditional domain generalization (DG) tackles this by including data from multiple source domains, which is impractical due to expensive data collection and annotation. Recent vision-language models like CLIP enable source-free domain generalization (SFDG) by using text prompts to simulate visual representations, reducing data demands. However, existing SFDG methods struggle with domain-specific confounders, limiting their generalization capabilities. To address this issue, we propose TDCRL (\textbf{T}ext-\textbf{D}riven \textbf{C}ausal \textbf{R}epresentation \textbf{L}earning), the first method to integrate causal inference into the SFDG setting. TDCRL operates in two steps: first, it employs data augmentation to generate style word vectors, combining them with class information to generate text embeddings to simulate visual representations; second, it trains a causal intervention network with a confounder dictionary to extract domain-invariant features. Grounded in causal learning, our approach offers a clear and effective mechanism to achieve robust, domain-invariant features, ensuring robust generalization. Extensive experiments on PACS, VLCS, OfficeHome, and DomainNet show state-of-the-art performance, proving TDCRL effectiveness in SFDG.

Mikel De Iturrate Reyzabal, Dionysios Malas, Shuai Wang et al.

We present a new approach for vision-based force estimation in Minimally Invasive Robotic Surgery based on frequency domain basis of motion of organs derived directly from video. Using internal movements generated by natural processes like breathing or the cardiac cycle, we infer the image-space basis of the motion on the frequency domain. As we are working with this representation, we discretize the problem to a limited amount of low-frequencies to build an image-space mechanical model of the environment. We use this pre-built model to define our force estimation problem as a dynamic constraint problem. We demonstrate that this method can estimate point contact forces reliably for silicone phantom and ex-vivo experiments, matching real readings from a force sensor. In addition, we perform qualitative experiments in which we synthesize coherent force textures from surgical videos over a certain region of interest selected by the user. Our method demonstrates good results for both quantitative and qualitative analysis, providing a good starting point for a purely vision-based method for surgical force estimation.

Huan Xu, Jinlin Wu, Guanglin Cao et al.

Ultrasonography has revolutionized non-invasive diagnostic methodologies, significantly enhancing patient outcomes across various medical domains. Despite its advancements, integrating ultrasound technology with robotic systems for automated scans presents challenges, including limited command understanding and dynamic execution capabilities. To address these challenges, this paper introduces a novel Ultrasound Embodied Intelligence system that synergistically combines ultrasound robots with large language models (LLMs) and domain-specific knowledge augmentation, enhancing ultrasound robots' intelligence and operational efficiency. Our approach employs a dual strategy: firstly, integrating LLMs with ultrasound robots to interpret doctors' verbal instructions into precise motion planning through a comprehensive understanding of ultrasound domain knowledge, including APIs and operational manuals; secondly, incorporating a dynamic execution mechanism, allowing for real-time adjustments to scanning plans based on patient movements or procedural errors. We demonstrate the effectiveness of our system through extensive experiments, including ablation studies and comparisons across various models, showcasing significant improvements in executing medical procedures from verbal commands. Our findings suggest that the proposed system improves the efficiency and quality of ultrasound scans and paves the way for further advancements in autonomous medical scanning technologies, with the potential to transform non-invasive diagnostics and streamline medical workflows.

Lukas Lindenroth, Danail Stoyanov, Kawal Rhode et al.

With soft robotics being increasingly employed in settings demanding high and controlled contact forces, recent research has demonstrated the use of soft robots to estimate or intrinsically sense forces without requiring external sensing mechanisms. Whilst this has mainly been shown in tendon-based continuum manipulators or deformable robots comprising of push-pull rod actuation, fluid drives still pose great challenges due to high actuation variability and nonlinear mechanical system responses. In this work we investigate the capabilities of a hydraulic, parallel soft robot to intrinsically sense and subsequently control contact forces. A comprehensive algorithm is derived for static, quasi-static and dynamic force sensing which relies on fluid volume and pressure information of the system. The algorithm is validated for a single degree-of-freedom soft fluidic actuator. Results indicate that axial forces acting on a single actuator can be estimated with an accuracy of 0.56 +- 0.66N within the validated range of 0 to 6N in a quasi-static configuration. The force sensing methodology is applied to force control in a single actuator as well as the coupled parallel robot. It can be seen that forces are accurately controllable for both systems, with the capability of controlling directional contact forces in case of the multi degree-of-freedom parallel soft robot.

Shuangyi Wang, Kehao Wang, Hongbin Liu et al.

Nasopharyngeal (NP) swab sampling is an effective approach for the diagnosis of coronavirus disease 2019 (COVID-19). Medical staffs carrying out the task of collecting NP specimens are in close contact with the suspected patient, thereby posing a high risk of cross-infection. We propose a low-cost miniature robot that can be easily assembled and remotely controlled. The system includes an active end-effector, a passive positioning arm, and a detachable swab gripper with integrated force sensing capability. The cost of the materials for building this robot is 55 USD and the total weight of the functional part is 0.23kg. The design of the force sensing swab gripper was justified using Finite Element (FE) modeling and the performances of the robot were validated with a simulation phantom and three pig noses. FE analysis indicated a 0.5mm magnitude displacement of the gripper's sensing beam, which meets the ideal detecting range of the optoelectronic sensor. Studies on both the phantom and the pig nose demonstrated the successful operation of the robot during the collection task. The average forces were found to be 0.35N and 0.85N, respectively. It is concluded that the proposed robot is promising and could be further developed to be used in vivo.

Lukas Lindenroth, Richard James Housden, Shuangyi Wang et al.

Objective: In this work we address limitations in state-of-the-art ultrasound robots by designing and integrating a novel soft robotic system for ultrasound imaging. It employs the inherent qualities of soft fluidic actuators to establish safe, adaptable interaction between ultrasound probe and patient. Methods: We acquire clinical data to determine the movement ranges and force levels required in prenatal foetal ultrasound imaging and design the soft robotic end-effector accordingly. We verify its mechanical characteristics, derive and validate a kinetostatic model and demonstrate controllability and imaging capabilities on an ultrasound phantom. Results: The soft robot exhibits the desired stiffness characteristics and is able to reach 100% of the required workspace when no external force is present, and 95% of the workspace when considering its compliance. The model can accurately predict the end-effector pose with a mean error of 1.18+/-0.29mm in position and 0.92+/-0.47deg in orientation. The derived controller is, with an average position error of 0.39mm, able to track a target pose efficiently without and with externally applied loads. Ultrasound images acquired with the system are of equally good quality compared to a manual sonographer scan. Conclusion: The system is able to withstand loads commonly applied during foetal ultrasound scans and remains controllable with a motion range similar to manual scanning. Significance: The proposed soft robot presents a safe, cost-effective solution to offloading sonographers in day-to-day scanning routines. The design and modelling paradigms are greatly generalizable and particularly suitable for designing soft robots for physical interaction tasks.

Shuangyi Wang, James Housden, Yohan Noh et al.

The development of robotic-assisted extracorporeal ultrasound systems has a long history and a number of projects have been proposed since the 1990s focusing on different technical aspects. These aim to resolve the deficiencies of on-site manual manipulation of hand-held ultrasound probes. This paper presents the recent ongoing developments of a series of bespoke robotic systems, including both single-arm and dual-arm versions, for a project known as intelligent Fetal Imaging and Diagnosis (iFIND). After a brief review of the development history of the extracorporeal ultrasound robotic system used for fetal and abdominal examinations, the specific aim of the iFIND robots, the design evolution, the implementation details of each version, and the initial clinical feedback of the iFIND robot series are presented. Based on the preliminary testing of these newly-proposed robots on 42 volunteers, the successful and re-liable working of the mechatronic systems were validated. Analysis of a participant questionnaire indicates a comfortable scanning experience for the volunteers and a good acceptance rate to being scanned by the robots.

Shan Luo, Wenxuan Mou, Kaspar Althoefer et al.

For humans, both the proprioception and touch sensing are highly utilized when performing haptic perception. However, most approaches in robotics use only either proprioceptive data or touch data in haptic object recognition. In this paper, we present a novel method named Iterative Closest Labeled Point (iCLAP) to link the kinesthetic cues and tactile patterns fundamentally and also introduce its extensions to recognize object shapes. In the training phase, the iCLAP first clusters the features of tactile readings into a codebook and assigns these features with distinct label numbers. A 4D point cloud of the object is then formed by taking the label numbers of the tactile features as an additional dimension to the 3D sensor positions; hence, the two sensing modalities are merged to achieve a synthesized perception of the touched object. Furthermore, we developed and validated hybrid fusion strategies, product based and weighted sum based, to combine decisions obtained from iCLAP and single sensing modalities. Extensive experimentation demonstrates a dramatic improvement of object recognition using the proposed methods and it shows great potential to enhance robot perception ability.

Shan Luo, Joao Bimbo, Ravinder Dahiya et al.

Touch sensing can help robots understand their sur- rounding environment, and in particular the objects they interact with. To this end, roboticists have, in the last few decades, developed several tactile sensing solutions, extensively reported in the literature. Research into interpreting the conveyed tactile information has also started to attract increasing attention in recent years. However, a comprehensive study on this topic is yet to be reported. In an effort to collect and summarize the major scientific achievements in the area, this survey extensively reviews current trends in robot tactile perception of object properties. Available tactile sensing technologies are briefly presented before an extensive review on tactile recognition of object properties. The object properties that are targeted by this review are shape, surface material and object pose. The role of touch sensing in combination with other sensing sources is also discussed. In this review, open issues are identified and future directions for applying tactile sensing in different tasks are suggested.

Shan Luo, Wenxuan Mou, Kaspar Althoefer et al.

Tactile data and kinesthetic cues are two important sensing sources in robot object recognition and are complementary to each other. In this paper, we propose a novel algorithm named Iterative Closest Labeled Point (iCLAP) to recognize objects using both tactile and kinesthetic information.The iCLAP first assigns different local tactile features with distinct label numbers. The label numbers of the tactile features together with their associated 3D positions form a 4D point cloud of the object. In this manner, the two sensing modalities are merged to form a synthesized perception of the touched object. To recognize an object, the partial 4D point cloud obtained from a number of touches iteratively matches with all the reference cloud models to identify the best fit. An extensive evaluation study with 20 real objects shows that our proposed iCLAP approach outperforms those using either of the separate sensing modalities, with a substantial recognition rate improvement of up to 18%.

Shan Luo, Leqi Zhu, Kaspar Althoefer et al.

This paper presents a successful application of deep learning for object recognition based on acoustic data. The shortcomings of previously employed approaches where handcrafted features describing the acoustic data are being used, include limiting the capability of the found representation to be widely applicable and facing the risk of capturing only insignificant characteristics for a task. In contrast, there is no need to define the feature representation format when using multilayer/deep learning architecture methods: features can be learned from raw sensor data without defining discriminative characteristics a-priori. In this paper, stacked denoising autoencoders are applied to train a deep learning model. Knocking each object in our test set 120 times with a marker pen to obtain the auditory data, thirty different objects were successfully classified in our experiment and each object was knocked 120 times by a marker pen to obtain the auditory data. By employing the proposed deep learning framework, a high accuracy of 91.50% was achieved. A traditional method using handcrafted features with a shallow classifier was taken as a benchmark and the attained recognition rate was only 58.22%. Interestingly, a recognition rate of 82.00% was achieved when using a shallow classifier with raw acoustic data as input. In addition, we could show that the time taken to classify one object using deep learning was far less (by a factor of more than 6) than utilizing the traditional method. It was also explored how different model parameters in our deep architecture affect the recognition performance.