Towards vision-based deep reinforcement learning for robotic motion control

This paper introduces a machine learning based system for controlling a robotic manipulator with visual perception only. The capability to autonomously learn robot controllers solely from raw-pixel images and without any prior knowledge of configuration is shown for the first time. We build upon the success of recent deep reinforcement learning and develop a system for learning target reaching with a three-joint robot manipulator using external visual observation. A Deep Q Network (DQN) was demonstrated to perform target reaching after training in simulation. Transferring the network to real hardware and real observation in a naive approach failed, but experiments show that the network works when replacing camera images with synthetic images.

Total energy shaping of a class of underactuated Port-Hamiltonian Systems using a new set of closed-loop potential shape variables

This paper proposes a method for designing set-point regulation controllers for a class of underactuated mechanical systems in Port-Hamiltonian System (PHS) form. A new set of potential shape variables in closed loop is proposed, which can replace the set of open loop shape variables-the configuration variables that appear in the kinetic energy. With this choice, the closed-loop potential energy contains free functions of the new variables. By expressing the regulation objective in terms of these new potential shape variables, the desired equilibrium can be assigned and there is freedom to reshape the potential energy to achieve performance whilst maintaining the PHS form in closed loop. This complements contemporary results in the literature, which preserve the open-loop shape variables. As a case study, we consider a robotic manipulator mounted on a flexible base and compensate for the motion of the base while positioning the end effector with respect to the ground reference. We compare the proposed control strategy with special cases that correspond to other energy shaping strategies previously proposed in the literature.

Tracking control of a class of Hamiltonian mechanical systems with disturbances

This paper presents a trajectory-tracking control strategy for a class of mechanical systems in Hamiltonian form. The class is characterised by a simplectic interconnection arising from the use of generalised coordinates and full actuation. The tracking error dynamic is modelled as a port-Hamiltonian Systems (PHS). The control action is designed to take the error dynamics into a desired closed-loop PHS characterised by a constant mass matrix and a potential energy with a minimum at the origin. A transformation of the momentum and a feedback control is exploited to obtain a constant generalised mass matrix in closed loop. The stability of the close-loop system is shown using the close-loop Hamiltonian as a Lyapunov function. The paper also considers the addition of integral action to design a robust controller that ensures tracking in spite of disturbances. As a case study, the proposed control design methodology is applied to a fully actuated robotic manipulator.

Development of a robust framework for an outdoor mobile manipulation UAV

There is a growing interest to autonomously collect or manipulate objects in remote or unknown environments, such as mountains, gullies, bush-land, or rough terrain. There are several limitations of conventional methods using manned or remotely controlled aircraft. The capability of small Unmanned Aerial Vehicles (UAV) used in parallel with robotic manipulators could overcome some of these limitations. By enabling the autonomous exploration of both naturally hazardous environments, or areas which are biologically, chemically, or radioactively contaminated, it is possible to collect samples and data from such environments without directly exposing personnel to such risks. This paper covers the design, integration, and initial testing of a framework for outdoor mobile manipulation UAV. The framework is designed to allow further integration and testing of complex control theories, with the capability to operate outdoors in unknown environments. The results obtained act as a reference for the effectiveness of the integrated sensors and low-level control methods used for the preliminary testing, as well as identifying the key technologies needed for the development of an outdoor capable system.

Cooperation using a robotic ad hoc network made from Bluetooth, JXTA, OSGi and other commercial off the shelf (COTS) products

Abstract - Mobile devices in the near future will need to collaborate to fulfill their function. Collaboration will be done by communication. We use a real world example of robotic soccer to come up with the necessary structures required for robotic communication. A review of related work is done and it is found no examples come close to providing a RANET. The robotic ad hoc network (RANET) we suggest uses existing structures pulled from the areas of wireless networks, peer to peer and software life-cycle management. Gaps are found in the existing structures so we describe how to extend some structures to satisfy the design. The RANET design supports robot cooperation by exchanging messages, discovering needed skills that other robots on the network may possess and the transfer of these skills. The network is built on top of a Bluetooth wireless network and uses JXTA to communicate and transfer skills. OSGi bundles form the skills that can be transferred. To test the nal design a reference implementation is done. Deficiencies in some third party software is found, specifically JXTA and JamVM and GNU Classpath. Lastly we look at how to fix the deciencies by porting the JXTA C implementation to the target robotic platform and potentially eliminating the TCP/IP layer, using UDP instead of TCP or using an adaptive TCP/IP stack. We also propose a future areas of investigation; how to seed the configuration for the Personal area network (PAN) Bluetooth protocol extension so a Bluetooth TCP/IP link is more quickly formed and using the STP to allow multi-hop messaging and transfer of skills.

Automated symbolic and numeric procedure for manipulator rigid-body dynamic significance analysis and simplification

This paper describes an automated procedure for analysing the significance of each of the many terms in the equations of motion for a serial-link robot manipulator. Significance analysis provides insight into the rigid-body dynamic effects that are significant locally or globally in the manipulator's state space. Deleting those terms that do not contribute significantly to the total joint torque can greatly reduce the computational burden for online control, and a Monte-Carlo style simulation is used to investigate the errors thus introduced. The procedures described are a hybrid of symbolic and numeric techniques, and can be readily implemented using standard computer algebra packages.

The implementation of a novel, bio-inspired, robotic security system

The implementation of a robotic security solution generally requires one algorithm to route the robot around the environment and another algorithm to perform anomaly detection. Solutions to the routing problem require the robot to have a good estimate of its own pose. We present a novel security system that uses metrics generated by the localisation algorithm to perform adaptive anomaly detection. The localisation algorithm is a vision-based SLAM solution called RatSLAM, based on mechanisms within the hippocampus. The anomaly detection algorithm is based on the mechanisms used by the immune system to identify threats to the body. The system is explored using data gathered within an unmodified office environment. It is shown that the algorithm successfully reacts to the presence of people and objects in areas where they are not usually present and is tolerised against the presence of people in environments that are usually dynamic.

Extending the dynamic range of robotic vision

Conventional cameras have limited dynamic range, and as a result vision-based robots cannot effectively view an environment made up of both sunny outdoor areas and darker indoor areas. This paper presents an approach to extend the effective dynamic range of a camera, achieved by changing the exposure level of the camera in real-time to form a sequence of images which collectively cover a wide range of radiance. Individual control algorithms for each image have been developed to maximize the viewable area across the sequence. Spatial discrepancies between images, caused by the moving robot, are improved by a real-time image registration process. The sequence is then combined by merging color and contour information. By integrating these techniques it becomes possible to operate a vision-based robot in wide radiance range scenes.

A new framework for force feedback teleoperation of robotic vehicles based on optical flow

This paper proposes the use of optical flow from a moving robot to provide force feedback to an operator's joystick to facilitate collision free teleoperation. Optic flow is measured by wide angle cameras on board the vehicle and used to generate a virtual environmental force that is reflected to the user through the joystick, as well as feeding back into the control of the vehicle. The coupling between optical flow (velocity) and force is modelled as an impedance - in this case an optical impedance. We show that the proposed control is dissipative and prevents the vehicle colliding with the environment as well as providing the operator with a natural feel for the remote environment. The paper focuses on applications to aerial robotics vehicles, however, the ideas apply directly to other force actuated vehicles such as submersibles or space vehicles, and the authors believe the approach has potential for control of terrestrial vehicles and even teleoperation of manipulators. Experimental results are provided for a simulated aerial robot in a virtual environment controlled by a haptic joystick.

Off-world robotic excavation for Large-scale habitat construction and resource extraction

This paper describes technologies we have developed to perform autonomous large-scale off-world excavation. A scale dragline excavator of size similar to that required for lunar excavation was made capable of autonomous control. Systems have been put in place to allow remote operation of the machine from anywhere in the world. Algorithms have been developed for complete autonomous digging and dumping of material taking into account machine and terrain constraints and regolith variability. Experimental results are presented showing the ability to autonomously excavate and move large amounts of regolith and accurately place it at a specified location.

DDX : A distributed software architecture for robotic systems

The Dynamic Data eXchange (DDX) is our third generation platform for building distributed robot controllers. DDX allows a coalition of programs to share data at run-time through an efficient shared memory mechanism managed by a store. Further, stores on multiple machines can be linked by means of a global catalog and data is moved between the stores on an as needed basis by multi-casting. Heterogeneous computer systems are handled. We describe the architecture of DDX and the standard clients we have developed that let us rapidly build complex control systems with minimal coding.

Performance tests for visual servo control systems, with application to partitioned approaches to visual servo control

Visual servoing has been a viable method of robot manipulator control for more than a decade. Initial developments involved positionbased visual servoing (PBVS), in which the control signal exists in Cartesian space. The younger method, image-based visual servoing (IBVS), has seen considerable development in recent years. PBVS and IBVS offer tradeoffs in performance, and neither can solve all tasks that may confront a robot. In response to these issues, several methods have been devised that partition the control scheme, allowing some motions to be performed in the manner of a PBVS system, while the remaining motions are performed using an IBVS approach. To date, there has been little research that explores the relative strengths and weaknesses of these methods. In this paper we present such an evaluation. We have chosen three recent visual servo approaches for evaluation in addition to the traditional PBVS and IBVS approaches. We posit a set of performance metrics that measure quantitatively the performance of a visual servo controller for a specific task. We then evaluate each of the candidate visual servo methods for four canonical tasks with simulations and with experiments in a robotic work cell.