Due to the high labor cost of physicians, it is difficult to collect a rich amount of manually-labeled medical images for developing learning-based computer-aided diagnosis (CADx) systems or segmentation algorithms. To tackle this issue, we reshape the image segmentation task as an image-to-image (I2I) translation problem and propose a retinal vascular segmentation network, which can achieve good cross-domain generalizability even with a small amount of training data. We devise primarily two components to facilitate this I2I-based segmentation method. The first is the constraints provided by the proposed gradient-vector-flow (GVF) loss, and, the second is a two-stage Unet (2Unet) generator with a skip connection. This configuration makes 2Unet's first-stage play a role similar to conventional Unet, but forces 2Unet's second stage to learn to be a refinement module. Extensive experiments show that by re-casting retinal vessel segmentation as an image-to-image translation problem, our I2I translator-based segmentation subnetwork achieves better cross-domain generalizability than existing segmentation methods. Our model, trained on one dataset, e.g., DRIVE, can produce segmentation results stably on datasets of other domains, e.g., CHASE-DB1, STARE, HRF, and DIARETDB1, even in low-shot circumstances.

We propose an adversarial training framework to simultaneously address the vessel segmentation and dirt/reflection removal problems in fundus photographs used for diabetic retinopathy diagnosis. This framework contains two primary subnetworks, each triggered by a set of loss terms, i.e., one for segmentation and the other for reconstruction. These two subnetworks act as inverse functions of each other so that they form an autoencoder framework with a 2-dimensional latent code, which can be a vessel segmentation mask after binarization. To further improve the segmentation and reconstruction performance, we devise a loss function based on gradient vector flow (GVF) and reorganize the generator network. Experimental results show that the proposed method has a good generalization capability. Trained on DRIVE’s training set, our model can produce segmentation and reconstruction-based artifact removal results stably on other datasets like CHASE-DB1 and STARE. The average F1-score of our segmentation results of DRIVE’s testing set reaches 0.7964, and the artifact-free reconstruction results can achieve an average PSNR of 24 dB.

Decentralized agent groups typically require complex mechanisms to accomplish coordinated tasks. In contrast, biological systems can achieve intelligent group behaviors with each agent performing simple sensing and actions. We summarize our recent papers on a biologically-inspired control framework for multi-agent tasks that is based on a simple and iterative control law. We theoretically analyze important aspects of this decentralized approach, such as the convergence and scalability, and further demonstrate how this approach applies to real-world applications with a diverse set of multi-agent applications. These results provide a deeper understanding of the contrast between centralized and decentralized algorithms in multi-agent tasks and autonomous robot control.

Modular robots can potentially assemble into a wide range of configurations to locomote in different environments. However, designing locomotion strategies for each configuration is often tedious and has generally relied on a priori known connection geometry. Here we present a framework for 2D modular robots made of square modules assembled with arbitrary geometry, which achieve collective and directed locomotion with no centralized controller. Individual modules communicate locally and provably achieve consensus in coordinating movement in a common travel direction. In experiments with simulations and hardware prototypes, we show that robots achieve effective locomotion, irrespective of the number of modules and their connectivity which can be highly asymmetric.

Biological systems achieve amazing adaptive behavior with local agents performing simple sensing and actions. Modular robots with similar properties can potentially achieve self-adaptation tasks robustly. Inspired by this principle, we present a generalized distributed consensus framework for self-adaptation tasks in modular robotics. We demonstrate that a variety of modular robotic systems and tasks can be formulated within such a framework, including (1) an adaptive column that can adapt to external force, (2) a modular gripper that can manipulate fragile objects, and (3) a modular tetrahedral robot that can locomote towards a light source. We also show that control algorithms derived from this framework are provably correct. In real robot experiments, we demonstrate that such a control scheme is robust towards real world sensing and actuation noise. This framework can potentially be applied to a wide range of distributed robotics applications.

In nature, animal groups achieve robustness and scalability with each individual executes a simple and adaptive strategy. Inspired by this phenomenon, we propose a decentralized control framework for modular robots to achieve coordinated and self-adaptive tasks with each modules performs simple distributed sensing and actuation. In this demonstration, we show that such a framework allows several different modular robotic systems to achieve self-adaptation tasks scalably and robustly, examples tasks include module-formed table and bridge that adapt to constantly-perturbed environment, a 3D relief display that renders sophisticated objects, and a tetrahedral robot that performs adaptive locomotion.

We present a modular robot design inspired by the creation of complex structures and functions in biology via deformation. Our design is based on the Tensegrity model of cellular structure, where active filaments within the cell contract and expand to control individual cell shape, and sheets of such cells undergo large-scale shape change through the cooperative action of connected cells. Such deformations play a role in many processes, e.g. early embryo shape change and lamprey locomotion. Modular robotic systems that replicate the basic deformable multicellular structure have the potential to quickly generate large-scale shape change and create dynamic shapes to achieve different global functions. Based on this principle, our design includes four different modular components: (1) active links, (2) passive links, (3) surface membranes, and (4) interfacing cubes. In hardware implementation, we show several self-deformable structures that can be generated from these components, including a self-deformable surface, expandable cube, terrain-adaptive bridge [C.-H. Yu et al., 2007]. We present experiments to demonstrate that such robotic structures are able to perform real time deformation to adapt to different environments. In simulation, we show that these components can be configured into a variety of bio-inspired robots, such as an amoeba-like robot and a tissue-inspired material. We argue that self-deformation is well-suited for dynamic and sensing-adaptive shape change in modular robotics.

Modular robots have the potential to achieve a wide range of applications by reconfiguring their shapes to perform different functions. This requires robust and scalable control algorithms that can form a wide range of user-specified shapes, including shapes that adapt to the environment. Here we present a decentralized algorithm for self-organizing of environmentally-adaptive shapes. We apply it to a chain-style modular robot, configured to form a flexible sheet structure. We show that the proposed algorithm is capable of achieving a wide class of environmentally-adaptive shapes, and the module control is simple, scalable, robust and provably correct. The algorithm is also self-maintaining: the shape automatically adapts if the environment changes. Finally, we present several applications which can be achieved within this framework via robot prototypes and simulations, such as a self-balancing table. In our experiments, we demonstrate the algorithm is highly responsive and robust in the face of real-world actuation and sensing noise.

Legged robots can, in principle, traverse a large variety of obstacles and terrains. In this paper, we describe a successful application of reinforcement learning to the problem of negotiating obstacles with a quadruped robot. Our algorithm is based on a two-level hierarchical decomposition of the task, in which the high-level controller selects the sequence of foot-placement positions, and the low-level controller generates the continuous motions to move each foot to the specified positions. The high-level controller uses an estimate of the value function to guide its search; this estimate is learned partially from supervised data. The low-level controller is obtained via policy search. We demonstrate that our robot can successfully climb over a variety of obstacles which were not seen at training time