Showing papers on "Physics engine published in 2019"

A Differentiable Physics Engine for Deep Learning in Robotics

Abstract: An important field in robotics is the optimization of controllers. Currently, robots are often treated as a black box in this optimization process, which is the reason why derivative-free optimization methods such as evolutionary algorithms or reinforcement learning are omnipresent. When gradient-based methods are used, models are kept small or rely on finite difference approximations for the Jacobian. This method quickly grows expensive with increasing numbers of parameters, such as found in deep learning. We propose the implementation of a modern physics engine, which can differentiate control parameters. This engine is implemented for both CPU and GPU. Firstly, this paper shows how such an engine speeds up the optimization process, even for small problems. Furthermore, it explains why this is an alternative approach to deep Q-learning, for using deep learning in robotics. Finally, we argue that this is a big step for deep learning in robotics, as it opens up new possibilities to optimize robots, both in hardware and software.

Propagation Networks for Model-Based Control Under Partial Observation

Abstract: There has been an increasing interest in learning dynamics simulators for model-based control. Compared with off-the-shelf physics engines, a learnable simulator can quickly adapt to unseen objects, scenes, and tasks. However, existing models like interaction networks only work for fully observable systems; they also only consider pairwise interactions within a single time step, both restricting their use in practical systems. We introduce Propagation Networks (PropNet), a differentiable, learnable dynamics model that handles partially observable scenarios and enables instantaneous propagation of signals beyond pairwise interactions. With these innovations, our propagation networks not only outperform current learnable physics engines in forward simulation, but also achieves superior performance on various control tasks. Compared with existing deep reinforcement learning algorithms, model-based control with propagation networks is more accurate, efficient, and generalizable to novel, partially observable scenes and tasks.

Interactive Differentiable Simulation

Abstract: Intelligent agents need a physical understanding of the world to predict the impact of their actions in the future. While learning-based models of the environment dynamics have contributed to significant improvements in sample efficiency compared to model-free reinforcement learning algorithms, they typically fail to generalize to system states beyond the training data, while often grounding their predictions on non-interpretable latent variables. We introduce Interactive Differentiable Simulation (IDS), a differentiable physics engine, that allows for efficient, accurate inference of physical properties of rigid-body systems. Integrated into deep learning architectures, our model is able to accomplish system identification using visual input, leading to an interpretable model of the world whose parameters have physical meaning. We present experiments showing automatic task-based robot design and parameter estimation for nonlinear dynamical systems by automatically calculating gradients in IDS. When integrated into an adaptive model-predictive control algorithm, our approach exhibits orders of magnitude improvements in sample efficiency over model-free reinforcement learning algorithms on challenging nonlinear control domains.

Physics Based Path Planning for Autonomous Tracked Vehicle in Challenging Terrain

Abstract: This paper describes a novel physics-based path planning architecture for autonomous navigation of tracked vehicles in rough terrain conditions. Unlike conventional path planning applications for smooth and structured environments, factors such as slip, slope of the terrain, robot actuator limitations, and dynamics of robot terrain interactions must be considered for rough terrain applications. The proposed path planning method consists of a hybrid planner/simulator, which takes into account all of the above factors by simulating the closed loop motion of the robot with a low-level controller on a realistic terrain model inside a physics engine. Once a feasible path to the goal is obtained, the same low-level closed loop controller is then used to execute the proposed path on the actual robot. The proposed architecture uses the D* Lite algorithm working on a 2D grid representation of the terrain as the high-level planner, Bullet as the physics engine and a hybrid automaton as the low-level closed loop controller. The proposed method is validated both in simulation and through experiments. Inferences based on the results from simulations and experiments show that the proposed planner is more effective in providing an optimal feasible path as compared to existing methodologies, demonstrating clear advantages for rough, unstructured terrain planning. Based on the results, possible improvements to the method are proposed for future work.

Combining Physical Simulators and Object-Based Networks for Control

Abstract: Physics engines play an important role in robot planning and control; however, many real-world control problems involve complex contact dynamics that cannot be characterized analytically. Most physics engines therefore employ approximations that lead to a loss in precision. In this paper, we propose a hybrid dynamics model, simulator-augmented interaction networks (SAIN), combining a physics engine with an object-based neural network for dynamics modeling. Compared with existing models that are purely analytical or purely data-driven, our hybrid model captures the dynamics of interacting objects in a more accurate and data-efficient manner. Experiments both in simulation and on a real robot suggest that it also leads to better performance when used in complex control tasks. Finally, we show that our model generalizes to novel environments with varying object shapes and materials.

Transferring Generalizable Motor Primitives From Simulation to Real World

Abstract: Reinforcement learning provides robots with an autonomous learning framework where a skill can be learned by exploration. Exploration in real world is, however, inherently unsafe and time consuming, and causes wear and tear. To address these, learning policies in simulation and then transferring them to physical systems has been proposed. In this letter, we propose a novel sample-efficient transfer approach, which is agnostic to the dynamics of a simulated system and combines it with incremental learning. Instead of transferring a single control policy, we transfer a generalizable contextual policy generated in simulation using one or few samples from real world to a target global model, which can generate policies across parameterized real-world situations. We studied the generalization capability of the incremental transfer framework using MuJoCo physics engine and KUKA LBR 4+. Experiments with ball-in-a-cup and basketball tasks demonstrated that the target model improved the generalization capability beyond the direct use of the source model indicating the effectiveness of the proposed framework. Experiments also indicated that the transfer capability depends on the generalization capability of the corresponding source model, similarity between source and target environment, and number of samples used for transferring.

Physics-as-Inverse-Graphics: Unsupervised Physical Parameter Estimation from Video

Abstract: We propose a model that is able to perform unsupervised physical parameter estimation of systems from video, where the differential equations governing the scene dynamics are known, but labeled states or objects are not available. Existing physical scene understanding methods require either object state supervision, or do not integrate with differentiable physics to learn interpretable system parameters and states. We address this problem through a physics-as-inverse-graphics approach that brings together vision-as-inverse-graphics and differentiable physics engines, enabling objects and explicit state and velocity representations to be discovered. This framework allows us to perform long term extrapolative video prediction, as well as vision-based model-predictive control. Our approach significantly outperforms related unsupervised methods in long-term future frame prediction of systems with interacting objects (such as ball-spring or 3-body gravitational systems), due to its ability to build dynamics into the model as an inductive bias. We further show the value of this tight vision-physics integration by demonstrating data-efficient learning of vision-actuated model-based control for a pendulum system. We also show that the controller's interpretability provides unique capabilities in goal-driven control and physical reasoning for zero-data adaptation.

Combining Coarse and Fine Physics for Manipulation using Parallel-in-Time Integration

Abstract: We present a method for fast and accurate physics-based predictions during non-prehensile manipulation planning and control. Given an initial state and a sequence of controls, the problem of predicting the resulting sequence of states is a key component of a variety of model-based planning and control algorithms. We propose combining a coarse (i.e. computationally cheap but not very accurate) predictive physics model, with a fine (i.e. computationally expensive but accurate) predictive physics model, to generate a hybrid model that is at the required speed and accuracy for a given manipulation task. Our approach is based on the Parareal algorithm, a parallel-in-time integration method used for computing numerical solutions for general systems of ordinary differential equations. We use Parareal to combine a coarse pushing model with an off-the-shelf physics engine to deliver physics-based predictions that are as accurate as the physics engine but runs in substantially less wall-clock time, thanks to Parareal being amenable to parallelization. We use these physics-based predictions in a model-predictive-control framework based on trajectory optimization, to plan pushing actions that avoid an obstacle and reach a goal location. We show that by combining the two physics models, we can achieve the same success rates as the planner that uses the off-the-shelf physics engine directly, but significantly faster. We present experiments in simulation and on a real robotic setup.

Physics-as-Inverse-Graphics: Joint Unsupervised Learning of Objects and Physics from Video.

Abstract: We propose a model that is able to perform unsupervised physical parameter estimation of systems from video, where the differential equations governing the scene dynamics are known, but labeled states or objects are not available. Existing physical scene understanding methods require either object state supervision, or do not integrate with differentiable physics to learn interpretable system parameters and states. We address this problem through a physics-as-inverse-graphics approach that brings together vision-as-inverse-graphics and differentiable physics engines, enabling objects and explicit state and velocity representations to be discovered. This framework allows us to perform long term extrapolative video prediction, as well as vision-based model-predictive control. Our approach significantly outperforms related unsupervised methods in long-term future frame prediction of systems with interacting objects (such as ball-spring or 3-body gravitational systems), due to its ability to build dynamics into the model as an inductive bias. We further show the value of this tight vision-physics integration by demonstrating data-efficient learning of vision-actuated model-based control for a pendulum system. We also show that the controller's interpretability provides unique capabilities in goal-driven control and physical reasoning for zero-data adaptation.

Haptically enabled simulation system for firearm shooting training

Abstract: Firearm shooting training is of importance in military and law enforcement training tasks. Traditional training usually uses actual firearms or modified bullets that are dangerous and expensive and difficult to evaluate the performance. Firearm training simulation systems provide risk-free alternatives. However, most existing simulation is visual-only, which lacks the immersion on the force feedback. In this paper, we proposed a new firearm training simulation system, which can provide more realistic training by incorporating physic effects on recoil and trigger pull weight. Dynamic, immersive, and repeatable training experiences while imposes no danger to trainees are provided in our system. The system consists of haptics, physics engine, and motion capture. These three components are carefully combined by developing the corresponding techniques of haptic force rendering, visuo-haptic integration, and synchronisation, physics-based dynamic simulation and motion analysis. Compared with existing systems, our training system has more complete functionalities that include visual firearm shooting, force generation, shooting reactions, result analysis and evaluation. Moreover, it is adaptable to off-the-shelf hardware and software packages and thus it can provide flexibility to system scalability and budget. To evaluate the proposed system, two demonstrations are conducted for users where the systems accuracy, immersion and usability are analysed. The results show the effectiveness of our physics-based shooting model and the proposed system on simulating different shooting scenarios.

Evolutionary Robotics Applied to Hexapod Locomotion: a Comparative Study of Simulation Techniques

Abstract: The Evolutionary Robotics (ER) process has been applied extensively to developing control programs to achieve locomotion in legged robots, as an automated alternative to the arduous task of manually creating control programs for such robots. The evolution of such controllers is typically performed in simulation by making use of a physics engine-based robotic simulator. Making use of such physics-based simulators does, however, have certain challenges associated with it, such as these simulators’ computational inefficiency, potential issues with lack of accuracy and the human effort required to construct such simulators. The current study therefore proposed and investigated an alternative method of simulation for a hexapod (six-legged) robot in the ER process, and directly compared this newly-proposed simulation method to traditional physics-based simulation. This alternative robotic simulator was built based solely on experimental data acquired directly from observing the behaviour of the robot. This data was used to construct a simulator for the robot based on Artificial Neural Networks (ANNs). To compare this novel simulation method to traditional physics simulation, the ANN-based simulators were used to evolve simple open-loop locomotion controllers for the robot in simulation. The real-world performance of these controllers was compared to that of controllers evolved in a more traditional physics-based simulator. The obtained results indicated that the use of ANN-based simulators produced controllers which could successfully perform the required locomotion task on the real-world robot. In addition, the controllers evolved using the ANN-based simulators allowed the real-world robot to move further than those evolved in the physics-based simulator and the ANN-based simulators were vastly more computationally efficient than the physics-based simulator. This study thus decisively indicated that ANN-based simulators offer a superior alternative to widely-used physics simulators in ER for the locomotion task considered.

Combining Physical Simulators and Object-Based Networks for Control

Abstract: Physics engines play an important role in robot planning and control; however, many real-world control problems involve complex contact dynamics that cannot be characterized analytically. Most physics engines therefore employ . approximations that lead to a loss in precision. In this paper, we propose a hybrid dynamics model, simulator-augmented interaction networks (SAIN), combining a physics engine with an object-based neural network for dynamics modeling. Compared with existing models that are purely analytical or purely data-driven, our hybrid model captures the dynamics of interacting objects in a more accurate and data-efficient manner.Experiments both in simulation and on a real robot suggest that it also leads to better performance when used in complex control tasks. Finally, we show that our model generalizes to novel environments with varying object shapes and materials.

Simulation of Realistic Particles with Bullet Physics Engine

Abstract: The traditional discrete element method (DEM) uses clumps to approximate realistic particles, which is computationally demanding when simulating many particles. In this paper, the Bullet physics engine is applied as an alternative to simulate realistic particles. Bullet was originally developed for computer games to simulate physical and mechanical processes that occur in the real world to produce realistic game experiences. Physics engines integrate a variety of techniques to simulate complex physical processes in games, such as rigid bodies (e.g., rocks, and soil particles), soft bodies (e.g., clothes), and their interactions. Therefore, physics engines have the capabilities to simulate realistic particles. This paper integrates three-dimensional laser scanner and Bullet to form a realistic particle simulation framework. The soil specimen collapse process is simulated to demonstrate the capability of the proposed framework to simulate realistic particles.

qiBullet, a Bullet-based simulator for the Pepper and NAO robots.

Abstract: The Pepper and NAO robots are widely used for in-store advertizing and education, but also as robotic platforms for research purposes. Their presence in the academic field is expressed through various publications, multiple collaborative projects, and by being the standard platforms of two different RoboCup leagues. Developing, gathering data and training humanoid robots can be tedious: iteratively repeating specific tasks can present risks for the robots, and some environments can be difficult to setup. Software tools allowing to simulate complex environments and the dynamics of robots can thus alleviate that problem, allowing to perform the aforementioned processes on virtual models. One current drawback of the Pepper and NAO platforms is the lack of a physically accurate simulation tool, allowing to test scenarios involving repetitive movements and contacts with the environment on a virtual robot. In this paper, we introduce the qiBullet simulation tool, using the Bullet physics engine to provide such a solution for the Pepper and NAO robots.

Gym-Ignition: Reproducible Robotic Simulations for Reinforcement Learning

Abstract: This paper presents Gym-Ignition, a new framework to create reproducible robotic environments for reinforcement learning research. It interfaces with the new generation of Gazebo, part of the Ignition Robotics suite, which provides three main improvements for reinforcement learning applications compared to the alternatives: 1) the modular architecture enables using the simulator as a C++ library, simplifying the interconnection with external software; 2) multiple physics and rendering engines are supported as plugins, simplifying their selection during the execution; 3) the new distributed simulation capability allows simulating complex scenarios while sharing the load on multiple workers and machines. The core of Gym-Ignition is a component that contains the Ignition Gazebo simulator and exposes a simple interface for its configuration and execution. We provide a Python package that allows developers to create robotic environments simulated in Ignition Gazebo. Environments expose the common OpenAI Gym interface, making them compatible out-of-the-box with third-party frameworks containing reinforcement learning algorithms. Simulations can be executed in both headless and GUI mode, the physics engine can run in accelerated mode, and instances can be parallelized. Furthermore, the Gym-Ignition software architecture provides abstraction of the Robot and the Task, making environments agnostic on the specific runtime. This abstraction allows their execution also in a real-time setting on actual robotic platforms, even if driven by different middlewares.

Physics Projection

Abstract: This study presents an approach termed physics projection, via which robots can learn about the physical world and predict the effects of their actions online and in an active manner. This approach employs three components: a robot, physical world model, and physics engine. The physics projection process involves a double loop structure comprising a real loop for learning the physical world model and an imaginary loop for a simulation search. Experiments were performed using the TurtleBot3 mobile robot and Unity graphics engine. The results effectively demonstrate that the robot can predict the effects of various actions performed by it under the given physical conditions, successfully execute the tasks of carrying a wine glass and a cup filled with water without dropping them or spilling their contents, and predict a catastrophic effect that could not be predicted by a human operator. The proposed method would contribute to enable robots to predict the effects of their actions and determine appropriate actions to perform in a dynamically changing physical world.

Learning Generalizable Physical Dynamics of 3D Rigid Objects

Abstract: Humans have a remarkable ability to predict the effect of physical interactions on the dynamics of objects. Endowing machines with this ability would allow important applications in areas like robotics and autonomous vehicles. In this work, we focus on predicting the dynamics of 3D rigid objects, in particular an object's final resting position and total rotation when subjected to an impulsive force. Different from previous work, our approach is capable of generalizing to unseen object shapes - an important requirement for real-world applications. To achieve this, we represent object shape as a 3D point cloud that is used as input to a neural network, making our approach agnostic to appearance variation. The design of our network is informed by an understanding of physical laws. We train our model with data from a physics engine that simulates the dynamics of a large number of shapes. Experiments show that we can accurately predict the resting position and total rotation for unseen object geometries.

Implementation of Generic Steering Algorithms for AI Agents in Computer Games

Abstract: This paper proposes a set of generic steering algorithms for autonomous AI agents along with the structure of the implementation of a movement layer designed to work with said algorithms. The algorithms are meant for further use in computer animation in computer games - they provide a smooth and realistic base for the animation of the agent’s movement and are designed to work with any graphic environment and physics engine, thus providing a solid, versatile layer of logic for computer game AI engines. Basic algorithms (called steering behaviors) based on dynamics have been thoroughly described, as well as some methods of combining the behaviors into more complex ones. Applications of the algorithms are demonstrated along with possible problems in their usage and the solutions to said problems. The paper also presents results of studies upon the behaviors within a closed, single-layered AI module consisting only out of a movement layer, thus removing the bias inflicted by pathfinding and decision making.

Online adaptation of uncertain models using neural network priors and partially observable planning

Abstract: One of the key challenges in realizing a robot that is capable of completing a variety of manipulation tasks in the real world is the need to utilize sufficiently compact and rich world models. If the assumed prediction model does not match real observations, planning systems are unable to perform properly. We propose a system that corrects the models based on information collected from the robot’s sensors. We encode prior experiences in a neural network to generate possible parameters of the models for a physics engine from real observations. An online POMDP solver is used to plan actions to complete the task while progressively validating and improving the models. We perform experiments in simulations and on a real robot. The results show that this approach appropriately clarifies observed environments, can handle dynamics with discontinuities, and with increasing domain complexity achieves a better success rate than baseline methods.

Simulator for disaster response robotics

Abstract: This chapter presents a simulator for disaster response robots based on the Choreonoid framework. Two physics engines and a graphics engine were developed and integrated into the framework. One physics engine enables robust contact-force computation among rigid bodies based on volumetric intersection and a relaxed constraint, whereas the other enables accurate and computationally efficient computation of machine–terrain interaction mechanics based on macro and microscopic approaches. The graphics engine allows simulating natural phenomena, such as rain, fire, and smoke, based on a particle system to resemble tough scenarios at disaster sites. In addition, wide-angle vision sensors, such as omnidirectional cameras and LIDAR sensors, can be simulated using multiple rendering screens. Overall, the simulator provides a tool for the efficient and safe development of disaster response robots.

A Practical Path Planning Method in Aircraft Assembly Process Simulation

Abstract: This paper proposes a method of solving aircraft assembly path planning by sampling search and implemented this method in virtual assembly application. This paper analyzes the advantages of two different sampling search methods, and finally selects the tree sampling search method. The Bullet physics engine is employed to implement obstacle detection. Finally, in the software platform, the simulation of the assembly process is completed based on the OpenSceneGraph visualization engine.

Aura projection for scalable real-time physics

Abstract: In this paper we propose a solution to delivering scalable realtime physics simulations. Although high performance computing simulations of physics related problems do exist, these are not realtime and do not model the real-time intricate interactions of rigid bodies for visual effect common in video games (favouring accuracy over real-time). As such, this paper presents the first approach to real-time delivery of scalable, commercial grade, video game quality physics. This is achieved by taking the physics engine out of the player's machine and deploying it across standard cloud based infrastructures. The simulation world is then divided into sections that are then allocated to servers. A server maintains the physics for all simulated objects in its section. Our contribution is the ability to maintain a scalable simulation by allowing object interaction across section boundaries using predictive migration techniques. We allow each object to project an aura that is used to determine object migration across servers to ensure seamless physics interactions between objects. The validity of our results is demonstrated through experimentation and benchmarking. Our approach allows player interaction at any point in real-time (influencing the simulation) in the same manner as any video game. We believe that this is the first successful demonstration of scalable real-time physics.

Accelerating Physics Engine Components with Embedded FPGAs

Abstract: In recent years there has been a steady increase in the use of physics engines, deployed in applications such as video games, scientific simulations, computer graphics and film productions. Their main purpose is to simulate the motions of objects based on real-world physics rules. As the complexity of the simulated scenes increases with the use of multiple objects and desirable effects, the computational cost of the physics-related calculations explodes. Typically, physics engines make use of the general-purpose computational capabilities of modern GPUs in order to take advantage of their massively parallel resources. In this paper, we consider the use of FPGAs to accelerate certain demanding components of the physics simulation pipeline aiming to provide better performing solutions at significantly lower energy cost. The results of our work demonstrate that by employing Zynq UltraScale+ devices featuring embedded ARM cores and FPGA fabric, we can accelerate physics computations of the popular Bullet library on highly demanding scenes up to 2.2x compared to high-end GPUs at a fraction of the energy required (up to 44x better energy efficiency).

Arc3 physics engine powered visual browser website design system and operating system

Abstract: Disclosed are the details about our computer system based software, and dedicated server system. The software is composed of a physics engine created with C++, and blueprinting techniques used to create an online web development platform. The physics engine acts as a base system on the computer that greatly enhances the features of a contemporary html, or java website. This system allows us to create more tools within a 3d metaverse, that is simple to use, and solves the 2d traditional website only problem in the internet. The invention has a wide variety of functions that allow 3d objects to be placed inside of the metaverse; customizing the site with a plethora of features including, color, material, images, and shape.

Mechanics–Based Virtual Prototyping of Robots with Deformable Bodies and Flexible Joints

Abstract: This paper describes a mechanics–based framework for virtual prototyping of soft robots, i.e. robots with deformable bodies and flexible joints. The framework builds on top of the screw theory, and uses geometrically exact nonlinear beam models for describing the behavior of deformable bodies, as well as the finite element method for space discretization. The computer implementation of this framework results in SimSOFT, a physics engine for soft robots. The capabilities of the framework are illustrated with one general example, an articulated chain of rigid and soft links connected through rigid and flexible joints. Furthermore, several case studies are shown for industrial and medical applications.