Joaquim Bento Cavalcante Neto

Evolving Plastic Neuromodulated Networks for Behavior Emergence of Autonomous Virtual Characters

This paper addresses the problem of generating natural behavior of autonomous virtual characters. Inspired by the fields of Embodied and Enactive Artificial Intelligence, we postulate that natural behavior is the result of a coupling between the agent and the world where it lives, which leads to a coherence between its actions and its surroundings. In this work, we present the tools that we have been using to study that idea: a controller based on a plastic neuromodulated neural network, which is capable of molding itself to received stimuli; and a simple novel method for genetic encoding of artificial neural networks. We show the capabilities of the controller in generating interesting foraging behavior of an autonomous virtual robot, and discuss the advantages of its emergent characteristics when compared with traditional approaches.

Emergence of Autonomous Behaviors of Virtual Characters through Simulated Reproduction

This paper addresses the problem of autonomous behaviors of virtual characters. We postulate that a behavior is regarded as autonomous when the actions performed by the agent result from the interaction between its internal dynamics and the environment, rather than being externally controlled. In this work, we argue that an autonomous behavior is an agent’s solution to a given problem, which is obtained through a process of self-organization of the dynamics of a system that is composed of the agent’s controller, its body and the environment. That process allows the emergence of complex behaviors without any description of actions or objectives. We show a technique capable of adapting an artificial neural network to consistently control virtual Khepera-like robots by means of simulated reproduction, with no measure of the robots’ fitness. All the robots are either male or female, and they are capable of evolving different kinds of behaviors according to their own characteristics, guided solely by the environment’s dynamics.

Cloth Simulation with Triangular Mesh Adaptivity

In the last decades, cloth animation has been the focus of much research, because of demands from the entertainment industry and from e-commerce. That type of animation is usually produced by means of physics simulations that are computationally expensive. Cloth folding during oscillations or due to contact with rigid objects often requires a very dense mesh when high curvatures are present. In those situations, the dynamics simulation will involve huge matrices and vectors. So, in the attempt to reduce costs, adaptive remeshing is frequently proposed. In this work, we investigate a remeshing approach during dynamics simulation of cloth. Mesh refinement is applied only to regions that need a fine level of detail. Our remeshing strategy refines the mesh in regions of high curvature and simplifies the mesh in regions of low curvature. No matter how regular and coarse the initial mesh is, our remeshing strategy produces meshes that are well adapted to the irregularities of the solid objects at every time step of the draping simulation. The fabric model consists of a triangular mesh and uses a spring-mass-damper system to compute the forces between particles, which are located at the meshs vertices. Collision detection depends on the arrangement of the cloth model and the objects in the scene. Although the tests show that, for comparable mesh sizes, the adaptive method does not always outperforms non-adaptive methods, the quality of the draping is much better when adaptive methods are used. Thus, adaptive methods can deliver comparable draping quality with fewer elements and less cost.

Requirements for Development of a Low Cost Portable Simulator for Shooting Skill Training

The use of simulators for shooting skill training is becoming more widespread with technological advances. However, this kind of training is still not accessible to many institutions because of its cost of acquisition, maintenance, portability and system use comprehensiveness. In this context, this paper proposes requirements for a viable model implementation of portable and low cost shooting simulator, while maintaining the level of immersion and focus on the training of a large number of simultaneous trainees. For the basis of the proposal, initially, there was a study on the market shooting training simulators, identifying key requirements. Therefore, with the guidance of experts in marksmanship training area, the requirements that serve the purpose of portability and low cost have been defined. Finally, a simulator was developed to validate the proposed model.

Generation of Emergent Navigation Behavior in Autonomous Agents Using Artificial Vision

In this work, we deal with the dynamics of the movements of autonomous agents, which are able to move in the environment using their own vision. For this, we apply the Continuous Time Recurrent Artificial Neural Network and the genetic encoding proposed in [1] [2]. However, we use a new sensorial description, which consists in captured images by a virtual camera, evolving an artificial visual cortex. The experiments show that the agents are able to navigate in the environment and to find the exit, in a non-programmed way, using only the visual data passed to the neural network. This has the flexibility to be applied in various environments, without displaying a forced tendency by a possible behavioral modeling as in other techniques.

New Method of Surface Reconstruction of Splat-Based Models

The surface splatting technique has several advantages in rendering models consisting of points, such as speed and quality of rendered images. After projecting all the splats on the plane of the image, it is necessary to reconstruct surfaces from several fragments generated in the previous step. Several heuristics make use of implicit information about the surface to determine which fragments constitute the same surface. In this work, a pre-processing performs the separation of surfaces by analyzing the normal vectors of the splats and their neighborhood relations. This step applies surface identifiers in splats in order to prevent fragments from different surfaces to be combined in the step of surface reconstruction.

Clipping of Splats Adapted to the Curvatures of the Edges

Surface splatting is a fast rendering technique for point-based models that usually delivers good-quality renderings. However, the intrinsic characteristic of this technique is the treatment of points as if they were discs. That requires special processing near sharp edges and corners in order to prevent rendering artifacts from appearing. In this work, we present a clever way of clipping splats near sharp edges and corners by a classification of neighboring splats, which belong to distinct intersecting surfaces. Those neighboring splats that take part in the clipping process are called \emph{clip partners}. Their concavity or convexity with respect to one another are determined, as well as their relation with the clipped splat. In our approach, there is no need of distance computations nor of other complex operations during the rendering process, since the classification can be performed offline during a pre-processing phase. Our approach is particularly important for virtual reality and game applications, where there is high probability of rendered objects to be viewed in a short distance by observers that navigate inside the virtual environment. The examples presented in this paper demonstrate the importance of correctly clipping splats for high quality rendering.