Abstract The correct and robust recognition of traffic signs is indispensable to self-driving vehicles and driver-assistant systems. In this work, we propose and evaluate two network architectures for multi-expert decision systems that we test on a challenging Traffic Sign Recognition Benchmark dataset. The decision systems implement individual experts in the form of deep convolutional neural networks (CNNs). A gating network CNN acts as final decision unit and learns which individual expert CNNs are likely to contribute to an overall meaningful classification of a traffic sign. The gating network then selects the outputs of those individual expert CNNs to be fused to form the final decision. In this work we study the advantages and challenges of the proposed multi-expert architectures that in comparison to other network architectures allow for parallel training of individual experts with reduced datasets. Under the challenging conditions introduced by the benchmark dataset, the demonstrated multi-expert decision systems achieve a recognition performance that is superior to those of humans: with an accuracy of 99.10%, when training experts with the complete dataset and 98.94%, when individual experts are only trained with 36% of the training samples. Overall, our approach ranked fourth on the list of the applied approaches proposed for the German traffic sign Recognition Benchmark (GTSRB) dataset.

This work addresses the instability in asynchronous data parallel optimization. It does so by introducing a novel distributed optimizer which is able to efficiently optimize a centralized model under communication constraints. The optimizer achieves this by pushing a normalized sequence of first-order gradients to a parameter server. This implies that the magnitude of a worker delta is smaller compared to an accumulated gradient, and provides a better direction towards a minimum compared to first-order gradients, which in turn also forces possible implicit momentum fluctuations to be more aligned since we make the assumption that all workers contribute towards a single minima. As a result, our approach mitigates the parameter staleness problem more effectively since staleness in asynchrony induces (implicit) momentum, and achieves a better convergence rate compared to other optimizers such as asynchronous \textsceasgd and \textscdynsgd, which we show empirically.

This paper develops a new method for coevolution, named Fitness-Diversity Driven Coevolution (FDDC). This approach builds on existing methods by a combination of a (predator-prey) Coevolutionary Genetic Algorithm (CGA) and novelty search. The innovation lies in replacing the absolute novelty measure with a relative one, called Fitness-Diversity. FDDC overcomes problems common in both CGAs (premature convergence and unbalanced coevolution) and in novelty search (construction of an archive). As a proof of principle, Spring Loaded Inverted Pendulums (SLIPs) are coevolved with 2D-terrains the SLIPs must learn to traverse.

This paper proposes an optical motion sensor aimed towards small robotic platforms. It incorporates a 20×20 pixel continuous-time CMOS silicon retina vision sensor with pixels that have local gain control and adapt to background lighting and a DSP microcontroller which computes the global optical flow from the sampled sensor output. The system allows the user to validate various motion algorithms suitable for the platform. Measurements are presented that show that the system can compute global 2D translational motion from complex natural scenes using the image interpolation algorithm at a sample rate of 1 kHz and for speeds up to ±1000 pixels/s using <5k instruction cycles per frame.

Roombots (RB) are self-reconfigurable modular robots designed to study robotic reconfiguration on a structured grid and adaptive locomotion off grid. One of the main goals of this platform is to create adaptive furniture inside living spaces such as homes or offices. To ease the control of RB modules in these environments, we propose a novel and more natural way of interaction with the RB modules on a RB grid, called the Natural Roombots User Interface. In our method, the user commands the RB modules using pointing gestures. The user's body is tracked using multiple Kinects. The user is also given real-time visual feedback of their physical actions and the state of the system via LED illumination electronics installed on both RB modules and the grid. We demonstrate how our interface can be used to efficiently control RB modules on simple point-to-point grid locomotion and conclude by discussing future extensions.

In this paper, a method is proposed for controlling a hopping monopod. It takes dynamics of the robot into account to have better nominal tracking of desired trajectories and more compliant environmental interactions at the same time. We have incorporated also natural dynamics of the robot into the system by using off-line gaits extracted from optimizations on energy. The main control loop consists of the projected inverse dynamics that generates actuator torques given desired trajectories and also a feedback loop designed and tuned specifically for the structure of the robot. A trajectory generator uses known optimal trajectories together with some stabilizing control laws that modify these trajectories to have better robustness in different situations. The average speed of the robot is also regulated by means of a self-organizing controller. We apply soft transitions in trajectories from phase to phase to avoid sharp actuator input profiles. Our method is successfully tested on a monopod hopper robot in simulation. It can handle slightly rough or sloped terrains while maintaining a given average speed. Simulation results suggest that our method is a promising candidate to control a real robot under construction.

Manipulation and transport of objects using mobile robotic platforms is a well studied field with several successful approaches. The main difficulty while using such platforms is the lack of adaptation capabilities to changes in the environment and the restriction to flat working areas. In this paper, we present a novel manipulation and transport framework using the self-reconfigurable modular robots Roombots to collaboratively carry arbitrarily shaped passive elements in a non-regular 3D environment equipped with passive connectors. A hierarchical planner based on the notion of virtual kinematic chain is used to generate collision-free and hardware-friendly paths as well as sequences of collaborative manipulations. To the best of our knowledge, this is the first example of manipulation of fully passive elements in an arbitrary 3D environment using mobile self-reconfigurable robots. The simulated results show that the planner is robust to arbitrary complex environments with randomly distributed connectors. In addition to simulation results, a proof of concept of the manipulation of one passive element with two real Roombots meta-modules is described.

The design of efficient locomotion gaits for robots with many degrees of freedom is challenging and time consuming even if optimization techniques are applied. Control parameters can be found through optimization in two ways: (i) through online optimization where the performance of a robot is measured while trying different control parameters on the actual hardware and (ii) through offline optimization by simulating the robot's behavior with the help of models of the robot and its environment. In this paper, we present a hybrid optimization method that combines the best properties of online and offline optimization to efficiently find locomotion gaits for arbitrary structures. In comparison to pure online optimization, both the number of experiments using robotic hardware as well as the total time required for finding efficient locomotion gaits get highly reduced by running the major part of the optimization process in simulation using a cluster of processors. The presented example shows that even for robots with a low number of degrees of freedom the time required for optimization can be reduced by a factor of 2.5 to 30, at least, depending on how extensive the search for optimized control parameters should be. Time for hardware experiments becomes minimal. More importantly, gaits that can possibly damage the robotic hardware can be filtered before being tried in hardware. Yet in contrast to pure offline optimization, we reach well matched behavior that allows a direct transfer of locomotion gaits from simulation to hardware. This is because through a meta-optimization we adapt not only the locomotion parameters but also the parameters for simulation models of the robot and environment allowing for a good matching of the robot behavior in simulation and hardware. We validate the proposed hybrid optimization method on a structure composed of two Roombots modules with a total number of six degrees of freedom. Roombots are self-reconfigurable modular robots that can form arbitrary structures with many degrees of freedom through an integrated active connection mechanism.

This paper presents the results of a study on the exploitation of compliance in structures made of self-reconfigurable modular robots - Roombots. This research was driven by the following three hypotheses: (1) compliance can improve locomotion performance; (2) different types of compliance will result in diverse locomotion behaviors; (3) control parameters optimized for a medium level of compliance will perform better for other values of compliance than parameters optimized for extremal compliance. Two types of in-series compliant elements were tested, with five different stiffness values for each of them, on a structure made of two Roombots modules. We ran dedicated on-line locomotion parameter optimizations for six different configurations and evaluated their performance for different stiffness values. Hypothesis 1 was confirmed for both types of compliant elements, with a peak of performance for an optimal level of compliance. The variety of locomotion strategies obtained for the different structures confirms hypothesis 2. Hypothesis 3 was only partially confirmed.

We present the design and evaluation of an iPad application that will be used to operate the modular robots “Roombots”. Roombots are the building blocks for adaptive pieces of furniture. The application allows a user to arrange adaptive furniture within a room. We conducted a user study with 24 participants to evaluate our approach and to freely explore people's interaction. Data suggests that the ability to move with the tablet leads to a better precision of the furniture arrangement. No significant difference has been observed between using the application through a virtual representation of the room in contrast to an augmented reality environment, even if participants mentioned in a post-study questionnaire their preference for the augmented condition. Users described the interface as intuitive and easy to use.

This paper presents our work towards a decentralized reconfiguration strategy for self-reconfiguring modular robots, assembling furniture-like structures from Roombots (RB) metamodules. We explore how reconfiguration by locomotion from a configuration A to a configuration B can be controlled in a distributed fashion. This is done using Roombots metamodules-two Roombots modules connected serially-that use broadcast signals, lookup tables of their movement space, assumptions about their neighborhood, and connections to a structured surface to collectively build desired structures without the need of a centralized planner.

We present a floating-gate synaptic circuit that updates its weight according to the Spike-Timing-Dependent Plasticity (STDP) rule. The weight (or floating-gate voltage) is updated only if the time difference between the pre- and post-synaptic spikes falls within a learning window. The update is implemented through tunneling and injection mechanisms which can be tuned for very long time constants up to seconds. The novelty of this circuit is that the tunneling and injection mechanisms are turned on only when the correlation of the pre and postsynaptic activity is significant. The additional benefit of this non-volatile technology is that synaptic weights can be stored locally on chip. We present experimental results that show the learning and normalization effects from the fabricated circuits.

We demonstrate the capabilies of a 1D motion detection chip in steering a car in a simulated racing game. The chip implemented in a 1.5 μm CMOS VLSI process, is comprised of 24 motion pixels that extracts optical flow in parallel. The local optical flow values go to a single-layer perceptron whose output is used to steer the car in the racing game. Because of the continuous-time operation of the motion detection chip, the computationally expensive task of generating a control signal for the car based on the visual scene is largely alleviated.

We present a Bluetooth scatternet protocol (SNP) that provides the user with a serial link to all connected members in a transparent wireless Bluetooth (BT) network. By using only local decision making we can reduce the overhead of our scatternet protocol dramatically. We show how our SNP software layer simplifies a variety of tasks like the synchronization of central pattern generator controllers for actuators, collecting sensory data and building modular robot structures. The whole BT software stack including our new scatternet layer is implemented on a single Bluetooth and memory chip. To verify and characterize the SNP we provide data from experiments using real hardware instead of software simulation. This gives a realistic overview of the scatternet performance showing higher order effects that are difficult to be simulated correctly and guarantees the correct function of the SNP in real world applications.

We describe a new motion detection circuit that extracts motion information based on a time-to-travel algorithm. The front-end photoreceptor adapts over 7 decades of background intensity and motion information can be extracted down to a contrast value of 2.5%. Results from the circuits which were fabricated in a 2-metal 2-poly 1.5um CMOS process, show that the motion information can be extracted over 2 decades of speed.