Business & Finance

Sharing the Load: Human-Robot Team Lifting Using Muscle Activity

Description
Seamless communication of desired motions and goals is essential for enabling effective physical human-robot collaboration. In such cases, muscle activity measured via surface electromyography (EMG) can provide insight into a person's intentions
Published
of 7
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  Sharing the Load: Human-Robot Team Lifting Using Muscle Activity Joseph DelPreto † and Daniela Rus †  Abstract —Seamless communication of desired motions andgoals is essential for enabling effective physical human-robotcollaboration. In such cases, muscle activity measured viasurface electromyography (EMG) can provide insight into aperson’s intentions while minimally distracting from the task.The presented system uses two muscle signals to create a controlframework for team lifting tasks in which a human and robotlift an object together. A continuous setpoint algorithm usesbiceps activity to estimate changes in the user’s hand height, andalso allows the user to explicitly adjust the robot by stiffeningor relaxing their arm. In addition to this pipeline, a neuralnetwork trained only on previous users classifies biceps andtriceps activity to detect up or down gestures on a rollingbasis; this enables finer control over the robot and expandsthe feasible workspace. The resulting system is evaluated by10 untrained subjects performing a variety of team lifting andassembly tasks with rigid and flexible objects. I. I NTRODUCTION Robots have the potential to provide humans with valuableassistance and greatly increase productivity, yet there isoften a communication barrier when trying to collaborate onphysical tasks. To facilitate natural interactions and efficientteamwork, an interface is needed that allows the robot toaccurately interpret the person’s intentions or commands ina way that minimally distracts from the ongoing task.Since a person naturally generates muscle activity duringphysical interactions, detecting these signals via surface elec-tromyography (sEMG) could provide valuable informationabout desired motion or stiffness. However, such signals aretypically dominated by noise and can be difficult to mapto limb motions due to the highly nonlinear relationshipsbetween electrical neuron activity, muscle activation, jointimpedances or torques, and limb dynamics. In addition, pre-dicting effective robot actions in response to known humanactions can be difficult and highly task-dependent.An approach to address these challenges could be blend-ing motion prediction from natural muscle activity with amuscle-based control interface for explicitly commandingadjustments. The method presented in this work creates sucha controller for team lifting tasks by using EMG signalsfrom the biceps and triceps as the only inputs, as depictedin Figure 1. A coarse estimation of the person’s upwardor downward motion is calculated from biceps activity andused to control the collaborating robot; the person can thenadjust their muscle activity to quickly increase or decreasethis setpoint. In addition, a plug-and-play neural network classifier detects up or down gestures at any time to offerfiner control and facilitate more complex lifting scenarios † MIT Distributed Robotics Lab, Cambridge, MA 02139 { delpreto, rus } @csail.mit.edu Fig. 1: A human and robot lift an object together, using muscle activityas the sole communication channel. Two pipelines process EMG signalsto estimate continuous height adjustments and detect up/down gestures. 1 where the person may want the robot to hold its end of theobject at a significantly different height than their own.This framework provides a human-robot communicationchannel that is embedded within the motions of the task itself. Muscle activity associated with the desired motionsis used to grant the person control over the robot and guideit towards providing effective assistance according to theirintentions and physical intuition. The system can then beapplied to a variety of lifting tasks.In particular, this work presents the following: •  an algorithm to continuously estimate a lifting setpointfrom biceps activity, roughly matching a person’s handheight while also providing a closed-loop control inter-face for quickly commanding coarse adjustments; •  a plug-and-play rolling classifier for detecting up ordown gestures from biceps and triceps activity, allowingthe user to explicitly command fine adjustments or movethe robot to targets farther from their own hand height; •  an end-to-end system integrating these pipelines to col-laboratively lift objects with a robot using only muscleactivity associated with the task; •  experiments with 10 subjects to evaluate the setpointand classification pipelines and to demonstrate collabo-rative assembly with rigid or flexible objects.II. R ELATED  W ORK This paper builds upon research exploring biosignals forrobot control and frameworks for human-robot collaboration.  A. Human-Robot Interaction There have been numerous approaches to understanding aperson’s intention and determining appropriate robot actions 1 Videos are available at http://people.csail.mit.edu/delpreto/icra2019  based on models of task and team dynamics. Studies haveexplored understanding the user’s perspective [1] and pre-dicting human intention or plan adaptation [2], [3]. Effectiveteam dynamics and cross-training have also been investigated[4], [5], [6], although the human-robot interface is often abottleneck for implementation [7].Physical interaction tasks have also been achieved usingmodalities such as vision, speech, force sensors, and gesture-tracking datagloves [8], [9], [10], [11], [12]. These may bedifficult to generalize to complex tasks though, and maybe hindered by interactions with the environment [13] orocclusions and ambient noise. Load-sharing policies havealso been developed for jointly manipulating objects withinsimulation [14], and planning methods are explored for safetyand efficiency [15], [16] or formation control [17].  B. Using Muscle Signals for Robot Control While models to predict appropriate actions given a per-ceived world state are powerful when they can be con-structed, directly conveying a user’s physical intentions canbe useful for interactive tasks. Many EMG devices havebeen developed for assistive robotics [18], including EMG-based exoskeletons for the hand [19], [20] or the upper-limb.Upper-limb exoskeletons have effectively used approachessuch as parameterized muscle models [21], fuzzy logiccontrollers [22], or impedance controllers [23]. Models canalso provide insights into muscle dynamics to facilitate thedevelopment of associated controllers [24], [25], [26], [27],[28]. Remote control or supervision of non-collaboratingrobots has also been explored via gesture-based control [29],[30], [31] or continuous trajectory estimation [32], [33]. Inaddition, muscle signals have been used to convey stiffnessinformation for dynamic physical collaboration tasks [34].Such studies have shown that EMG can yield effectivehuman-robot interfaces, but also demonstrate associated chal-lenges such as noise, variance between users, and complexmuscle dynamics. Examples of addressing such challengesinclude redundant switching models [33] or leveraging thehuman within an exoskeleton control loop [35].III. E XPERIMENTAL  S ETUP AND  S YSTEM  D ESIGN An experimental setup and a closed-loop system weredesigned to explore team lifting scenarios and evaluateperformance. The principal components and information floware illustrated in Figure 2.  A. Team Lifting Tasks Two pillars of LEDs indicate target lifting heights or cuetraining gestures. During lifting tasks, the user lifts an objectand controls the robot to achieve the desired height while theLEDs are lit, then lowers the object and robot back to thetable when they turn off. A dumbbell weighing up to 10lbis used for non-collaborative trials, while a shared object isused for trials with physical robot interaction.In addition to these structured lifting tasks, collaborativeassembly tasks are performed in which the user and robotinstall an object on a base structure. As seen in Figure 5, Fig. 2: The system consists of an experiment controller that coordinatesthe paradigm, EMG acquisition, two pipelines for computing a robotpose, the Baxter robot, and a human subject. Visual and/or physicalfeedback closes the loop. a variety of rigid and flexible objects are used. The usercan command the robot through any trajectory they deemappropriate based on the task and the object’s properties.  B. Experiment Controller and Robot  A centralized control program coordinates the experimen-tal paradigms. It controls event timing, decides target heights,and controls the cue LEDs. It also relays commands fromthe EMG subsystem to the robot and sends event triggers tothe EMG subsystem. A humanoid Rethink Robotics Baxterrobot was used for these experiments. C. EMG Hardware and Data Acquisition Muscle signals are acquired from the right Biceps Brachiishort head and the right Triceps Brachii long head usingdifferential pairs of adhesive 24mm Covidien electrodes.MyoWare Muscle Sensors provide amplified raw and filteredsignals, which are then sampled at 1kHz by an NI USB-6216data acquisition device and processed within Simulink 2017b.The setpoint algorithm uses the biceps signal, while theclassification pipeline uses the biceps and triceps signals.  D. Subject Selection A total of 10 subjects participated in the experiments (90%male, 80% right-handed). No previous experience usingEMG interfaces was required, and subjects were not screenedbased on EMG signals. All subjects provided written consentfor the study, which was approved by MIT’s Committee onthe Use of Humans as Experimental Subjects.IV. C ONTINUOUS  S ETPOINT  E STIMATION The setpoint algorithm aims to estimate changes in theperson’s hand height while also creating a task-based controlinterface. Rather than model the muscle and limb dynamicsto attempt an accurate prediction of pose and appropriaterobot response, it leverages the person’s ability to closethe loop based on visual and physical feedback. The robotroughly lifts to the same height as the person using only theirnatural muscle activity, then the user can consciously controlthe robot by adjusting their muscle activity.  Algorithm 1  Setpoint Control Algorithm S IGNAL  P ROCESSING 1:  raw ← amplified biceps EMG signal from MyoWare board, sampled at 1kHz 2:  filtered ← band-pass filter 5-400Hz 3:  envelope ← rectify, low-pass filter 5Hz, amplify by 1.5 4:  envelopeScaled ← normalize using MVC-based  G normalization U PDATE  B UFFERS 5:  Circular  baselineBuffer  and  enabledBuffer ← envelopeScaled 6:  Non-circular  changesBuffer ← envelopeScaled D ETERMINE  E NABLE  B IT 7:  enabled ← mean ( enabledBuffer )  > L enable 8:  if   enabled  then E STIMATE AND  F ILTER  S ETPOINT  A DJUSTMENTS 9:  if   changesBuffer  is full  then 10:  rawChange ← integrate [ changesBuffer − baselineBuffer ] 11:  if  | rawChange |  < L change  or  ( ≥  75%  of   actionHistory   = 0 and  majority of those  =  sign ( rawChange ) )  then 12:  rawChange ← 0 13:  end if  14:  setpointChange ← rawChange × G setpoint 15:  robotHeight ← robotHeight  +  setpointChange 16:  Circular  actionHistory  ← sign ( rawChange ) 17:  end if  18:  end if  Algorithm 1 outlines the pipeline, which processes a singleEMG channel from the biceps to produce two outputs: acontinuously adjusted robot height, and whether the personis engaged in the lifting task. Figure 3 shows sample results.  A. Signal Processing The amplified raw biceps signal from the MyoWare pro-cessing board is conditioned and processed in software toextract a signal envelope. A bandpass filter preserves theuseful frequency content of the EMG signal [36] whileremoving high-frequency noise and low-frequency offsetsor motion artifacts. The signal envelope is then detectedto indicate muscle activation levels. Example filtered andenvelope-detected signals are shown in Figure 3.  B. Parameterized Setpoint Algorithm The central principle of the setpoint algorithm is to focuson short-timescale changes in muscle activation around along-timescale baseline level of activation. By incrementallyupdating the setpoint based on relative changes in muscleactivity instead of mapping muscle activations to poses, thealgorithm is more robust to EMG variation across usersand time. In addition, changes in activation are intuitivelyinformative for lifting tasks since moving a weight betweentwo poses requires a spike in torque.There are 8 main parameters, which will be introducedthroughout this section then optimized in the next section.The algorithm first applies a normalization gain G normalization  based on maximum voluntary contraction(MVC) to the detected EMG envelope. This signal thenpopulates a circular buffer of duration  D baseline  and ashorter noncircular buffer of duration  D integration . Themean of the long buffer represents a baseline level of activation, which is subtracted from the shorter buffer Fig. 3: These sample traces were recorded while the human-robot teamlifted an object to 7 target heights. The pipeline filters, amplifies, andenvelope-detects the EMG signal (top), computes a rolling baseline todetermine whether the controller should be enabled (middle), and usesvariations around the baseline to adjust the robot setpoint (bottom). Thetargets (magenta) were achieved. TABLE I: Setpoint Algorithm Parameters Optimization Values UsedParameter TuningType  Bounds Result (mean ± SD) D baseline  One-time [0.25, 5.00]s 2.38s 2.40s D integration  One-time [0.05, 0.20]s 0.10s 0.10s D changes  One-time [0.20, 2.00]s 1.90s 1.90s D enable  One-time [0.38, 2.25]s 2.00s 1.00s L change  One-time – – 0.01 L enable  Per-Subject [0.01, 0.15] 0.15 0.28 ± 0.27 G normalization  Per-Subject – – 11.50 ± 10.48 G setpoint  Per-Subject [0.20, 20.00] 0.75 0.38 ± 0.20 to yield changes around the baseline. These changes areintegrated when the short buffer fills. If the result is above athreshold level  L change , it is amplified by a gain  G setpoint to adjust the robot’s setpoint.Before applying this change to the robot, however, thecomputed adjustments are slightly filtered. A rolling historyis stored spanning  D changes  seconds of whether the com-puted setpoint adjustments were positive, negative, or zero.A new setpoint adjustment is only applied if its sign agreeswith the majority of the stored history (ignoring zeros) or if atleast three-quarters of the history is zero. This filtering helpssmooth the robot motion, which can be especially desirablewhen the user moves rapidly or when muscle signals becomeerratic due to fatigue.To determine whether the controller should be enabled, arolling buffer of the normalized envelope spanning  D enable seconds is compared to a threshold  L enable . If this indicatesthe user is not lifting, the robot waits in a neutral position. C. Parameter Optimization A summary of the parameters required by the model isshown in Table I. Most of the parameters are held constantfor all users, and were optimized based on initial open-loopexperiments where a user was cued to lift a dumbbell toeach of 7 target heights. A genetic algorithm was used forall parameters except  G normalization  and  L change , whichwere set based on MVC and manual preference, respectively.Using the Matlab 2017b Global Optimization Toolbox, thealgorithm was run for 66 generations with a uniform creationfunction, rank scaling, scattered crossover fraction 0.8, elite  ratio 0.05, and stochastic uniform selection. The objectivefunction simulated the setpoint algorithm with new parame-ters, then extracted the root mean square (RMS) error of thecomputed setpoints at the end of each trial.Table I presents the results along with the values ultimatelyused during online experiments. The computed durations arereasonable given the cadence of a typical lifting task; thebaseline buffer spans a few seconds, the history of changesallows a reversal of lifting direction approximately once persecond, and the integration buffer is short yet long enoughto smooth spurious EMG fluctuations. Note that thresholdsand gains act on normalized signals.The three subject-specific parameters are based on brief calibration routines. The subject is cued to repeatedly tenseand relax their arm, then to lift a dumbbell to 7 targetheights without robot feedback.  G normalization  and  L enable are computed by comparing relaxation and contraction seg-ments, then  G setpoint  is computed by simulating the setpointalgorithm and minimizing its error at each target.  D. Human-in-the-Loop Control Although the model is optimized by minimizing positionerror between the user and robot, the system is not designedto track the human’s position accurately in online trials; theuser may not want the robot to mirror their own position. In-stead, the algorithm provides a coarse height estimation andcreates a framework with which the user can communicatedesired motions. Biceps are naturally used to lift an object,then further tensing or relaxing the muscle will further raiseor lower the robot. By using the task’s primary muscle ina way that is consistent with the task, the system aims toreduce necessary training and behavioral modification.Conceptually, the algorithm achieves this by leveragingthe independence between joint position and stiffness. Aperson can lift an object using mainly their biceps, butthen significantly change their muscle activation withoutmoving the object by co-activating their antagonistic triceps.The controller effectively repurposes this stiffness degree of freedom as a communication channel to the robot.V. R OLLING  G ESTURE  C LASSIFICATION While the setpoint algorithm estimates changes in theuser’s hand height and enables active control over the robot,it can be difficult to achieve fine-grained or persistent heightadjustments. In addition, commanding robot positions thatare significantly higher than the user’s hand may be tiring.To address these cases, a classification pipeline was imple-mented that operates in parallel with the setpoint algorithm.It continuously classifies EMG envelopes from the bicepsand triceps to detect up and down gestures; the robot thenbriefly moves slowly in the desired direction. This pipeline isalso plug-and-play, only trained on data from prior subjects,so new users can immediately control the robot via gestures.  A. Training Gestures and Feature Extraction Training data was collected by cueing up or down gesturesduring specified time windows using the LED pillars. As Fig. 4: EMG signals are segmented and processed into gesture trainingdata by normalizing, centering, shifting down to 0, and downsampling.The center columns represent 286 trials from 6 subjects, with meantraces in bold and one standard deviation shaded on each side. Syntheticaugmentation examples are not included. illustrated in Figure 1, an up gesture consists of two brief upward hand motions while a down gesture consists of asingle brief downward hand motion. As users become moreexperienced with the system, they can learn to produce therequired muscle activations while minimizing object motion.Subjects were instructed to match their gesture duration tothe LEDs, which remained lit for 1s. Figure 4 visualizescollected EMG signals and extracted training data, demon-strating some common characteristics but also significantvariability between gesture examples and across subjects. 1) Segmentation and Signal Processing:  Envelopes of biceps and triceps muscle activity are acquired from theMyoWare processing boards then normalized based on MVC.From each trial, one labeled gesture segment and one base-line segment are extracted according to the cue LEDs.The segmented muscle signals are smoothed by a movingmean with duration 75ms, and downsampled to 50Hz. Theyare then independently shifted so their minimum values are at0 to reduce the impact of signal drift, inter-subject variations,and the weight or height of the object. 2) Data Augmentation:  To achieve reliable predictions ona rolling basis despite training on time-locked examples, adata augmentation approach can be employed [29]. Eachextracted training segment is first centered; up gestures arecentered to the average location of two prominent bicepspeaks, while down gestures are centered to a single promi-nent triceps peak. Four positively labeled copies of eachgesture are then synthesized by randomly shifting left andright between 0 and 100ms. Four negatively labeled copiesare also synthesized by shifting farther left and right between150 and 450ms. Together, these encourage the network toprefer gestures centered in its classification window withina certain tolerance. For each srcinal baseline segment, tworandomly shifted copies are also synthesized.In addition to augmenting via time shifts, synthetic exam-ples were generated based on magnitude. While the signalsare normalized via MVC, there can still be variation acrosssubjects, object weights, or gesture speeds. After synthesiz-ing time-shifted examples, every example in the corpus iscopied and scaled by a random factor between 0.6 and 1.4.As a result of this augmentation, each gesture segmentyields 10 positive and 8 negative examples while eachbaseline segment yields 6 negative examples. The center 1.5sof each example is then extracted, and the two EMG channelsare concatenated to yield a 150-element feature vector.  Fig. 5: Various team lifting and assembly tasks were performed. Column(a) uses the setpoint algorithm with cued target heights in open-loop,separated closed-loop, and interactive closed-loop scenarios. Column (b)performs assembly with rigid and flexible objects. Column (c) extendsthe system to use stiffness and additional degrees of freedom.  B. Neural Network Training and Online Classification These labeled vectors are used to train a feed-forwardneural network using the Pattern Recognition functionalityof Matlab’s Neural Network Toolbox (2017b). The network has a single hidden layer of size 50 using a hyperbolictangent sigmoid activation function, and an output layer of size 3 using a softmax activation function. The three outputsare used to indicate whether the segment was classified asbaseline, a down gesture, or an up gesture.For each online experiment, a new classifier was trainedusing data from previous subjects. Streaming EMG envelopesare then normalized, smoothed, downsampled, and usedto populate two 1.5s rolling buffers in Simulink (2017b).With each new sample, the buffers are shifted down to 0,concatenated, and classified. Network outputs are slightlyfiltered to reduce spurious predictions; a rolling buffer of 40 classifications (800ms) is maintained, and a final gestureprediction is declared if its mode occurs at least 4 times andis on average within 40ms of the buffer’s center.If an up or down gesture is detected, the robot moves inthat direction at 3cm/s. It stops when the opposite gesture isdetected, 5s have elapsed without another gesture detected,or the robot’s height limits are reached.VI. E XPERIMENTAL  R ESULTS AND  D ISCUSSION Experiments were conducted with 10 subjects to evaluatethe algorithms and system efficacy. 1  A. Setpoint Algorithm To evaluate the setpoint algorithm, users lifted an object toa sequence of 7 increasing heights while the system estimatedan appropriate robot position. This was done in an open-loop scenario without robot motion, then in two closed-loopscenarios: one where the robot moved but did not interactwith the user, and one where the human and robot jointlylifted a rigid object. Figure 5a shows the task setups. Targetswere equally spaced between 10.0cm and 50.0cm above a76.5cm table. All 10 subjects participated in these scenarios,totalling 20 open-loop sequences (excluding 1 per subjectused for calibration), 22 separated closed-loop sequences,and 39 collaborative sequences.Figure 6a illustrates the results. Across all targets, themean and standard deviation of the RMS setpoint error Fig. 6: For each target lifting height (green), the achieved robot setpointsare aggregated across all users. All trials used the setpoint algorithm,but only those in (b) used gestures. There is significantly less errorwith closed-loop feedback, and using gestures further increases accuracywhile reaching higher targets. during the last 1.0s of each trial for the open-loop, sepa-rated closed-loop, and collaborative closed-loop cases were14.1 ± 15.2cm, 10.0 ± 11.1cm, and 9.0 ± 16.0cm, respec-tively. Users visually compared the object or the robot’sfingers to the LED targets, so some variability is expected.Each closed-loop case was significantly more accurate thanthe open-loop case (  p <  0 . 01  for each one), with littledifference between the two closed-loop cases (  p <  0 . 49 ).This suggests that including closed-loop feedback, eithervisual or physical, allows the human to significantly improveperformance by adjusting muscle activity to control the robot.The object was also level during the final 1.0s of collab-orative trials, with a pitch of -1.4 ◦ ± 5.4 ◦ . Together with theincreased accuracy between open-loop and closed-loop cases,this indicates that users successfully leveraged arm stiffnessto control the robot without affecting their own pose. 1) Assembly Tasks:  Two assembly tasks illustrated inFigure 5b were also performed. The human and robot jointlylifted an object, waited while the human manipulated a basestructure with their free hand, then lowered the object ontothe structure. This was done with a rigid object (8 subjects,41 trials) and a flexible rubber sheet (4 subjects, 20 trials).The rigid-object assembly lasted 16.4 ± 8.1s, during whichthe object pitch averaged -0.3 ◦ ± 5.3 ◦ . The flexible-objectassembly averaged 8.8 ± 3.8s. Assembly was successful inboth cases, demonstrating applicability of the system to basicteam lifting tasks with a variety of object material properties.  B. Gesture Classification To evaluate whether the gesture classification pipelinefacilitates finer control and higher targets, users were cued togesture the robot towards 3 targets from 62.5cm to 82.5cmabove the table while keeping their elbow at approximately90 ◦ . This was designed to mimic lifting a flexible carbonfiber sheet to a specific angle before pressing it against avertical mold. Figure 6b visualizes the results. Across all 3targets, spanning 38 trials from 5 subjects, the mean RMSerror during the final 1.0s of each trial was 6.8 ± 7.6cm.Users achieved lower error than in the separated closed-loopcase without gestures (  p <  0 . 05 ). And while the setpoint
Search
Similar documents
View more...
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks
SAVE OUR EARTH

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!

x