text stringlengths 128 2.05k |
|---|
These discoveries reveal that the MPFA improves foraging rates when the swarm size or the number of targets are scaled up. This is not only because of the simple intuitive reduction in travel time, but also because of the significant improvement in avoiding collisions. Overall, the MPFA has better performance as the sw... |
In the future work, we will discover the trends on the evolved seven parameters and confirm the above hypothesis for the random, partially clustered and clustered resource distributions. In addition, we will consider the cost of deploying multiply nests and evolve the optimized number of nests for different resource di... |
ACKNOWLEDGMENT We gratefully acknowledge members of the Moses Biological Computation Lab for their assistance with the multiple-place foraging swarm robotics project. Thanks to Antonio Griego for developing the CPFA algorithm in ARGoS. Thanks to Matthew Fricke for discussing the issues in the experiments. |
# Source: arxiv 1612.08126 # Title: Brain-Swarm Interface (BSI): Controlling a Swarm of Robots with Brain and Eye Signals from an EEG Headset # Sections: all # Downloaded: 2026-03-03T02:01:35.429433+00:00 |
Brain-Swarm Interface (BSI): Controlling a Swarm of Robots with Brain and Eye Signals from an EEG Headset Abstract This work presents a novel marriage of Swarm Robotics and Brain Computer Interface technology to produce an interface which connects a user to a swarm of robots. The proposed interface enables the user to ... |
INTRODUCTION In this paper we present a new brain machine interface for a human user to control a swarm of robots, which we call a Brain-Swarm Interface (BSI). The BSI uses an off-the-shelf Electroencephalogram (EEG) headset to record brain and muscle activity from the user’s scalp. We use both signals from brain neuro... |
Brain Computer Interfaces hold great promise for enabling people with various forms of disabilities, from restricted motion due to injury or old age, to severe disabilities like the ALS (Locked-in syndrome), Tetraplegia, and paralysis. Several works have investigated using BCIs for controlling prosthetics |
, and for medical rehabilitation. If implemented effectively, BCI technology may allow people with disabilities the power to manipulate their environment, and to move themselves within their environment |
. Researchers have also developed brain interfaces for controlling single mobile robot platforms. For example Bin He et al. have demonstrated 3D control of a quadcopter, and Tim Bredtl et al. |
have remotely teleoperated a UAV, both using motor imagery BCI. C J Bell et al. have controlled a humanoid with a non invasive BCI using P300 signals. However, brain computer interfaces for controlling swarms have received little attention in the literature. |
Whereas the motivation for a BCI operated prosthetic or wheelchair is evident, the applications for a brain swarm interface may be less obvious. We envision several applications for this technology. Firstly, people who are mobility-impaired may use a swarm of robots to manipulate their environment using a brain-swarm i... |
The brain is arguably the most complex organ in the human body. Brain activity occurs mainly in the chemical and electromagnetic domain. Different areas of the brain are responsible for different thoughts and actions. The EEG is one of the few ways in which brain activity can be monitored in a non-invasive manner. An E... |
and utilize the ”artifacts” movements, which are usually discarded in BCI systems, as control input to modulate the position and size of a robotic swarm. A survey of other hBCI systems, which do not include application to robotic swarms, can be found in |
One of the key components of a typical BCI paradigm is the machine learning and pattern recognition pipeline. A survey by F Lotte et al. |
surveys the various types of classifiers that have been used in the literature. In our work, we adopt a Hidden Markov Model (HMM) technique to estimate the user’s thought state, as the dynamic nature of the technique is well-suited to online estimation and control input generation. We use outputs from an Emotiv Epoc EE... |
The rest of this paper is organized as follows. In Section II we introduce notation and formally state the problem. In Section III we describe the signal processing and control components that go into our system. Section |
IV presents the results of simulation and hardware experiments, and conclusions are given in Section II PROBLEM FORMULATION Let the position of the swarm be described by the vector [MATH] where [MATH] denotes the number of agents in an [MATH] dimensional Euclidean space. The position of the [MATH] individual at time [M... |
[MATH] . Suppose the user has an intended trajectory for the swarm, which we denote [MATH] . When thinking about this trajectory, the user’s EEG headset produces signals at time |
[MATH] , denoted by [MATH] , where [MATH] is the number of sensors on the EEG headset. Let [MATH] be a vector of control parameters of the system given by [MATH] where |
[MATH] [MATH] [MATH] denote the attraction and repulsion gains, respectively, which are used to control the size of the swarm, and [MATH] denotes the displacement vector for controlling the movement of the swarm. We state our problem as follows. |
Problem 1 (Brain-Swarm Interface) Design a signal processing pipeline to determine the swarm control parameters [MATH] from the EEG signals [MATH] , so that the swarm trajectory approaches the user’s intended swarm trajectory, |
[MATH] To control the swarm, we adapt a potential field based swarm controller of a type that is common in controlling swarms of ground robots |
With this type of controller, the system dynamics are given by [EQUATION] where [MATH] and [MATH] is a function which depends on pairwise interactions between agents |
[MATH] and [MATH] , and has parameters [MATH] [MATH] , and [MATH] In our proposed solution, we first train an HMM, and use it to determine the “thought state” of the user [MATH] , where for example, [MATH] indicates “aggregation” and [MATH] indicates “dispersion.” Then we map this though state [MATH] to values for |
[MATH] to control the aggregation/dispersion of the swarm. Secondly, we design an eye movement classifier which takes EEG signals |
[MATH] , where [MATH] with [MATH] , as input to determine the user’s intended motion of the swarm, producing a control parameter [MATH] for the swarm. This pipeline is shown graphically in Fig. , and its components are described in more detail in the following section. |
III SYSTEM DESCRIPTION In the previous section we stated the problem, and introduced the main elements of our solution as shown in Fig . The following paragraphs will describe the HMM, eye movement detection, and formation control strategies in detail. |
III-A Decoding Thoughts Using an HMM Based on a survey by F. Lotte et al. we adopted Hidden Markov Models to train and classify EEG data. Previously, researchers have used HMMs for BCIs. Pfurtscheller et al. |
used an HMM to classify EEG data, and HMMs have been used in conjunction with other techniques , as well. Our novel HMM implementation uses performance metrics observations and maps them to discrete thoughts. We observe that each thought will corresponds to different signatures of these metrics. However, these signatur... |
Instead we implement a training phase to train the HMM to detect the users though signatures and transition probabilities. Using standard terminology and notation for HMMs, |
, here we will describe our system in detail. III-A HMM: An Overview A Hidden Markov Model is a joint probabilistic model of a collection of discrete random variables [MATH] described by: |
[EQUATION] Many algorithms exist for both learning the transition probabilities and observation probabilities of such a system from data, and determining a likely sequence of states hidden [MATH] from data. The observations of these states [MATH] can be discrete or continuous. |
The main components of the model from Equation are: [MATH] which is the initial state probability distribution and is represented by [MATH] where [MATH] |
[MATH] is the state transition probability represented by the matrix [MATH] where [MATH] [MATH] is the observation probability represented by the gaussian distribution |
[MATH] with means [MATH] and covariance matrix [MATH] The model can be completely described by parameters [MATH] . The initial phase involves learning these parameters [MATH] to generate the model using training data consisting of observations of the expected state space. We employ the Baum-Welch algorithm (a version o... |
We are use a two state HMM which represents two distinct thoughts of the user, so [MATH] in our case. The observation space consists of the EEG output [MATH] . In this case, the signal [MATH] is derived from performance metrics provided by the manufacturer of the EEG headset, Emotiv. Emotiv provides six metrics: “Engag... |
III-A The Training Phase with Baum-Welch Algorithm Training data consisting of the three metrics mentioned before is recorded in a single trial. During the training period the user repeats two thoughts through a pre-defined switching sequence. We enforced this thought sequence using a timed slide presentation which the... |
The EM Algorithm is a two step iterative process (Expectation followed by Maximization), which can be expressed in the single expression |
[EQUATION] where [MATH] is the log likelihood function, [MATH] is the unobserved state, [MATH] is the unknown parameters of the model, and [MATH] is the observed variable. The expectation step can be summarized by calculating the following quantities: |
[EQUATION] where [MATH] is known as the forward variable, and is the probability of ending in state [MATH] and seeing the partial observations [MATH] given the model parameters [MATH] and |
[EQUATION] where [MATH] is known as the backward variable, and is the probability of observing partial sequences [MATH] given the model parameters and state at time [MATH] . With [MATH] and [MATH] |
we can compute [EQUATION] where [MATH] is the probability of being in state [MATH] given the model parameters and observations, as well as |
[EQUATION] where [MATH] is the probability of being in state [MATH] at time [MATH] and state [MATH] at time [MATH] Using the above defined quantities in the expectation step we can estimate the parameters [MATH] in the maximization step of our system as follows : |
[EQUATION] Naturally, for this procedure to start we need [MATH] which is the set of initial model parameters from which recursion begins according to Eqn. ( ). We adopt a K-means clustering approach to initialize the model parameters, specifically the mean matrix |
[MATH] . We used a two class K-means approach with the observations [MATH] as inputs, which give us the centroids of the two classes that we used to initialize [MATH] . The other parameters of [MATH] are initialized randomly. We stop the iterative process when we observe only minute changes (order of [MATH] ) in the es... |
A typical training signal from our experiments is shown in Fig. 3(a) and the resulting state sequence after training is shown in Fig. 3(b) . The user visited the two states (thoughts) twice each during the training period. From Fig. 3(b) one can see that the Baum-Welch algorithm detects the switching sequence between t... |
III-A Online Estimation with the Forward Algorithm After the model parameters [MATH] have been estimated we can employ the HMM for online estimation of the state. However, we cannot immediately proceed to online estimation. First, we have to relate the HMM states back to the original thoughts. Specifically, we do not k... |
Online estimation of state is now a straightforward application of the Forward algorithm for HMMs using the learn parameters from the training phase. This allows us to find the most likely sequence of thought states on line as a streaming signal arrives from the EEG. All we need for this phase is to calculate the value... |
from Eqn. for the current time [MATH] for all the states [MATH] and determine the most probable state at that time. So we can describe the control output from the HMM at time [MATH] given by |
[MATH] as [EQUATION] The output [MATH] is used to determine the the control parameters [MATH] for aggregation and dispersion of the swarm, as described in Sec. III-C below. |
III-B Tracking Eye Movements In traditional EEG research, Eye Movement signals are considered as artifacts and are removed. In contrast, we use these signals as inputs for our system to command the direction of travel for the robots. There are various available methods in the literature for detecting and tracking eye m... |
. These methods can be broadly categorized into (i) Contact based tracking which offer high accuracy and sophistication, (ii) Non-contact based optical tracking methods which measure relative positioning remotely with sensors such as cameras, and (iii) measuring surface electrical potentials from skin, also known as El... |
III-B EOG : An Overview The human eye can be modeled as an electrical dipole whose axis is roughly collinear to the axis of the human eye. The electrical dipole rotates with the rotation of the eye causing small differences (in microvolts) between the electrical potential at the skin surface depending on eye position. ... |
EOG typically uses exclusive electrodes around the eyes to measure movements. But our electrode positions are fixed so we employ the four closest electrodes to the eyes: ‘AF3’,‘AF4’,‘F7’ and ‘F8’, according to the 10-20 EEG sensor placement system, as shown in the diagram in Fig. . Previous methods to detect eye motion... |
The spatio-temporal signals from these electrodes near the eyes can be described by [MATH] where [MATH] denoting the electrodes used. We first normalize the signal by subtracting its mean for each electrode to center the signals about zero. This can be described by |
[EQUATION] where [MATH] denotes the number of samples used for the baseline removal. III-B Horizontal Eye Movement Detection Electrodes ’F7’ and ’F8’ are chosen for horizontal eye movement detection as they are the farthest apart in the horizontal plane while being closest to the eyes. Our algorithm for decoding horizo... |
Algorithm 1 Horizontal Eye Movement Detection 1: Remove Baseline with [MATH] samples. 2: Window the data with window size [MATH] |
[EQUATION] where [MATH] represents the window number. We use [MATH] samples corresponding to 1 second of data with no overlap. 3: Apply [MATH] |
order 4 Hz low pass Butterworth filter to the windowed data to isolate the eye movement signals. 4: Subtract the resulting signals from both electrodes |
[EQUATION] 5: Detect peaks and troughs with threshold magnitude of [MATH] and minimum seperation of [MATH] samples in [MATH] 6: Assign Peaks to eye movements to the left [MATH] |
and troughs to eye movements to the right [MATH] III-B Vertical Eye movement Detection Electrodes ’AF3’ and ’AF4’ are chosen for vertical eye movement detection. The method used is different from horizontal eye movement detection since we do not have any electrode below the eyes to detect the dipoles in the vertical pl... |
and described in Algorithm . In Fig , the green ellipses indicate the signal for vertical eye movement, and the blue ellipses indicate horizontal movement. The red ellipse represents blinks which are filtered out. |
Algorithm 2 Vertical Eye Movement Detection 1: Remove Baseline with [MATH] samples. 2: Window the data similar to Equation 13 3: Same as step 3 in Algorithm 1. |
4: Add the signals from both the electrodes to get [EQUATION] 5: Find peaks and troughs with minimum seperation of [MATH] and a signal level of [MATH] [MATH] from |
[MATH] 6: Assign Peaks to upward eye movements [MATH] and troughs to downward eye movements [MATH] It should be noted that both the algorithms above also filter out blinks, which look quite similar to the eye movements (see Figs. |
and ). In the case of horizontal eye movements, the F7 and F8 electrodes both record blinks with almost equal magnitude since they are located approximately at the same distance from the eyes. Hence the signal |
[MATH] is automatically devoid of blinks (Eqn 15 ), as evident from Fig . For the vertical eye movement detection, we introduce an upper threshold to filter out the blinks, which typically are much larger than eye movement signals, as can be seen in Fig. . Hence, our algorithm effectively discards blinks and measures o... |
Finally, after the left-right and up-down signals have been extracted form the user’s eye movements, these signals are used to control the left-right and forward-backward motion of the robot swarm through the control parameter [MATH] , as described below in Sec. III-C |
III-C Formation Control We described the general form of the potential field based formation controller above in Sec II . We refer the reader to |
for details and proofs about stability and convergence properties. Here we describe how we control the size and motion of the swarm through the parameters [MATH] |
III-C Controlling Size Recall that the system is described by the two dimensional state space equation for the [MATH] agent [EQUATION] |
We let the interaction between robots [MATH] and [MATH] be given by [EQUATION] Where [MATH] is the radius of the robot. We can see that the left term provides the attracting field, and the right the repelling field. The [MATH] term introduces a safety region around the robots so collision can be avoided. |
There is an equilibrium inter-robot distance for this system, in which attraction and repulsion forces balance. Let that equilibrium distance be denoted [MATH] , so that |
[MATH] . This [MATH] is governed by the attraction and repulsion gains [MATH] and [MATH] respectively. In our method we vary the gains to achieve different equilibrium formation. We map the two state output from the HMM to two distinct sets of gains in order to achieve the aggregation and dispersion of the swarm. |
III-C Controlling size and motion Now to control the motion of the swarm we rely on the output from the eye movement detection which gives us four possible motion commands: Forward, Backward, Left, and Right. We use these command to assign values to the vector [MATH] , which drives every robot in the swarm in the same ... |
IV SIMULATIONS AND HARDWARE EXPERIMENTS To demonstrate our brain-swarm interface, we developed a simulation environment in Matlab. We chose a section of the Boston University campus, with a rectangular path around a campus building, as shown in |
. The path is divided into 4 edges and the swarm has to be driven starting from the left of edge 1 and end on the top of edge 4 following a clockwise motion. At edge 3 (purple path) due to the narrow passageway, the user has to make the swarm aggregate into a tighter swarm by switching thoughts, while in edges 1, 2 and... |
For the training phase, the user switched between two thoughts at least twice over a period of 60 seconds, during which the EEG signals were recorded and fed into the Baum-Welch Algorithm (Fig. ) to get the model parameters as described in Sec. III-A . The thoughts used for simulations and experiment were distinct and ... |
The simulated swarm consisted of 128 point sized holonomic robots. The attraction gain [MATH] was fixed at [MATH] and the repulsion gain [MATH] was calculated according to [MATH] , where [MATH] is the estimated state sequence from ( 12 ). In our [MATH] state HMM case according to the previous formula the user’s thought... |
The results of the simulation exercise is summarized in Figs. and Fig. shows the time history of eye movement detection and the mental thought estimation during the motion of the swarm along the 4 legs of the path. It can be seen that the thought estimation remains mostly in the disperse state during legs 1, 2, and 4, ... |
For the hardware experiments we used the m3pi platform with an Mbed controller for mobile swarming robots, and Zigbee radios for communication. The experiments were carried out in an environment with an Optitrack motion capture system to track the motion of the robots (see Fig. ). The control parameters used were |
[MATH] and [MATH] for aggregation, and [MATH] and [MATH] for dispersion. These two wheeled differential drive robots receive individual motor speeds as control inputs from the computer. A proportional point-offset controller is used to generate the motor speeds from the potential function controller (the details of whi... |
). The control commands for the robots were computed off board the robots, and set to the robots over Zigbee at an update rate of 30 Hz. Due to computational and hardware complexities the computations were divided among 3 computers (one for Optitrack data acquisition, one for controller implementation, and another othe... |
The experimental results are summarized in Figs. 11 and 12 . Fig. 11 shows the time history of eye movement detection and the thought state estimation throughout the motion of the swarm along the 4 legs. From Fig. |
which shows the path of the robots and their centroid, we can see the eye movement detection is successful in steering the swarm. The color map for the thought state estimation again shows that the HMM is able to reliably determine the user’s intention. The system remains in the disperse state with high confidence duri... |
The biggest challenge in this work is in the integration of this complex system with interacting hardware, communication, software, and human components. We used holonomic dynamics during simulation whereas The M3pis are nonholonomic robots with inefficient actuation and communication. In addition, it was quite a menta... |
CONCLUSIONS In this paper we propose and successfully demonstrate an online Brain Swarm Interface to control a swarm of ground vehicles in simulation and experiments using off-the-shelf hardware. We integrate a variety of engineering and scientific techniques in neuroscience, signal processing, machine learning, contro... |
ACKNOWLEDGMENTS We thank Alyssa Pierson for her help in developing the point offset controller to control individual M3pi robots. |
# Source: arxiv 1612.09474 # Title: A generative angular model of protein structure evolution # Sections: all # Downloaded: 2026-03-03T01:59:20.935927+00:00 |
A generative angular model of protein structure evolution Abstract Recently described stochastic models of protein evolution have demonstrated that the inclusion of structural information in addition to amino acid sequences leads to a more reliable estimation of evolutionary parameters. We present a generative, evoluti... |
Abstract This supplement contains further details on the methods employed in the paper and additional results. Keywords: Evolution; Protein structure; Probabilistic model; Directional statistics. |
Introduction Recently, several studies (Challis and Schmidler, 2012 ; Herman et al. 2014 have proposed joint stochastic models of evolution which take into account simultaneous alignment of protein sequence and structure. These studies point out the limitations of earlier non-probabilistic methods, which often rely on ... |
We present a generative evolutionary model, ETDBN (Evolutionary Torus Dynamic Bayesian Network) for pairs of homologous proteins. ETDBN captures dependencies between sequence and structure evolution, accounts for alignment uncertainty, and models the local dependencies between aligned sites. This extends the models pre... |
ETDBN is motivated by the non-evolutionary TorusDBN model (Boomsma et al. 2008 . TorusDBN is a first-order Hidden Markov Model (HMM) that represents a single protein structure as a sequence of [MATH] dihedral angle pairs, which are modelled using continuous bivariate angular distributions (Frellsen et al. 2012 . Likewi... |
Each amino acid in a peptide chain is covalently bonded to the next via a chemical bond referred to as a peptide bond . Peptide bonds have a partial double bond nature, that results in a planar configuration of atoms in space. This configuration allows the protein backbone structure to be largely described in terms of ... |
The evolution of dihedral angles in ETDBN is modelled using a novel stochastic diffusion process developed in García-Portugués et al. 2017 . In addition to this, a coupling is introduced such that an amino acid change can lead to a jump in dihedral angles and a change in diffusion process, allowing us to capture change... |
Given that naturally occurring protein sequences are evolutionarily related, it is only natural to consider models that take such evolution into account. Two or more homologous proteins will share a common ancestor, which leads to underlying tree-like dependencies. These dependencies manifest themselves most noticeably... |
Failing to account for evolutionary dependencies can lead to false conclusions (Felsenstein, 1985 , whereas accounting for evolutionary dependencies allows information from homologous proteins to be incorporated in a principled manner. This can lead to more accurate inferences, such as the prediction of a protein struc... |
Parameters of ETDBN, particularly those governing the amino acid and dihedral angle evolutionary processes, were learnt during a training phase using 1200 homologous protein pairs from the HOMSTRAD database (Mizuguchi et al. 1998 . This resulted in a realistic prior distribution over proteins compared to previous stoch... |
Evolutionary model 2.1 Overview ETDBN is a dynamic Bayesian network model of local protein sequence and structure evolution along a pair of aligned homologous proteins [MATH] and [MATH] . ETDBN can be can be viewed as an HMM (see Figure ). Each hidden node of the HMM, corresponding to an aligned position, adopts an evo... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.