Motion Generation for Humanoid Robots with ... - Brown University

More documents

Recommendations

Info

primitive behavior database. Shepard’s interpolation [17]is used. The generated motion data are sent back to theplanner. The actions are then sequenced and sent to thehardware controller(s) to accomplish the robot's intentions.Figure 6: Embedding and clustering of first motion streamusing spatio-temporal Isomap. Lines show temporallyadjacent segments.100 random coordinates with the bounding boxmethod were interpolated to obtain new joint spacetrajectories for each primitive behavior [17]. The reachingbehavior and return to home behavior are interpolated(Figure 7).Figure 4. Structure of the search mechanism for generatingnew motions from stored behaviors.6 Validation6.1 ExperimentISAC is driven by a human operator to reach to 5random points (A, B, C, D, E) on a table, a limitedworking space, as shown in Figure 5.Figure 5: ISAC is taught to reach 5 points on a tableThe motion data stream consisted of 1241 frames,which contained only one reaching behavior. The streamwas segmented using the Kinematic Centroid Method [8]resulting in 10 segments. Each segment was normalizedto 100 frames. The parameters for the K-nearestneighborhood, common-temporal (c ctn ) and adjacenttemporal neighborhood (c atn ) constants for spatio-temporalIsomap were selected (*). The structure was embedded to3 dimensions. The segments that were similar came closerafter applying spatio-temporal Isomap (Figure 6; eachsegment in the embedding is numbered). The primitivefeature groups were produced through clustering. Thecloser segments were clustered in the embedded spacewithin 0.1 of the diagonal of the embedding bounding boxto derive the primitive behaviors [4]. The clusteringresulted in two derived primitive behaviors: reaching andreturning to home positions.(*) K –nearest neighbors = 3, c ctn =100 and c atn = 1.Figure 7: Results from interpolating selected action units.Each plot shows trajectories in Cartesian coordinates.(Left) reaching, (Right) returning to home.Spatio-temporal Isomap was applied a second time(**). Sweep-and-prune clustering was then used in theembedded space with a 0.1 threshold value to derive themeta-level behaviors (Figure 8). The segments from thesame primitive feature groups became proximal after thefirst application of Spatio-temporal Isomap. After thesecond application of spatio-temporal, the primitivestypically performed in sequence were collapsed (Figure 9).Thus segments were collapsed from the correspondingprimitive feature groups in the sequence into the higherlevelbehaviors. The meta-level embedding resulted in thesame number of clusters as in the primitive embedding.This convergence of the embeddings indicates there is nohigher-level structure in the motion greater than theprimitive level. Thus, we consider the vocabulary has onemeta-level behavior that encompasses the two primitivebehaviors.Figure 8: Embedding and clustering of second applicationof Spatio-temporal Isomap. Proximal motion segments aregrouped to form behaviors.(**) K-nearest neighbors 4, c ctn =100, c atn = 4.
Figure 9: The transitions between the segments that derivemeta-level behavior for each reaching motion. Linesindicate transitions between actions.6.2 ApplicationA human using a finger to point can produce ademonstration for ISAC, who attends to the objects in itsworkspace and generates reaching motions to thoseobjects. The application begins with a speech cue from thehuman, which directs the robot's attention to an object. Todirect ISAC's attention to unknown position of the object,the human tells ISAC to find the new the location of theobject such as “Reach-to”. The Human Agent sends thisintention text to the Self Agent (SA), and activates theHuman Finger Agent (HFA) inside the Human Agent andparses the name of the object. The HFA finds a pointedfinger to fixate on the object. Next, the Head Agent isactivated to find the pointed finger place and cameraangles information is forwarded to the Sensory EgoSphere,which returns the coordinates of the object. Based on theseintentions, the CEC uses procedures in LTM to retrieve themotion data to accomplish the robot's reaching intention.The desired motion data sequence is sent back to the CEC,and then sent to the Arm Agent to perform the reachingtask.Figure 10: Schematic representation of the systemcommunication for the demonstration of Reaching theNew Point7 Related WorkAutonomous control of humanoid robots has been atopic of significant research. Preprogramming and teleoperationremain common methods for most applications,such as lift trucks driven by robotic operators [5].However, these approaches are quite tedious for complextasks and environments. Huber and Grupen [6] havepresented hybrid architecture for autonomous controlgiven a manually constructed control basis.Consequently, their approach to control is dependent onthe quality of the design and construction of the controlbasis, whereas the automated derivation we used [8]leverages the structure of humans.Other methods for automatically deriving behaviorsfrom human motion are not always suitable forautonomous control. Bregler [1] presented automaticallydeveloping groups of behaviors in the form of movemesfrom image sequences of a moving human. Complexmotion can be described by sequencing the movemesgenerated for each limb, but indexing the movemes forcoordinated motion generation is not obvious. Li et al.[19] and Kovar et al. [10] presented work for buildinglinear dynamical systems and directed graphs from motioncapture. For generating motion for control, however, thesesystems require a constraint optimization to be applied onstructures that may not be parsimonious.Human motion has been used as a basis forcontrolling humanoid robots. Ijspeert et al. [7] presentedan approach for learning non-linear dynamical systemswith attractor properties from motion capture. Theirapproach is useful for perturbation-robust humanoidcontrol, but is restricted to deriving a single class ofmotion.Brooks and Stein [2] developed an integratedphysical system including vision, sound input and output,and skillful manipulation, which are all controlled by acontinuously operating parallel communication. The goalwas to enable the resulting system to learn to "think" bybuilding on its bodily experiences to accomplishprogressively more abstract tasks [2]. ISAC’s motionlearning system is similar. In our work, vision, speechrecognition, short-term memory, self-agent and behaviorsstored in long-term are all operating parallel in a uniquearchitecture in order to control the humanoid robotautonomously.8 ConclusionWe have described an approach for generating newmotions from derived behaviors. We were able to derivethe behavior vocabularies based on the Spatio-temporalIsomap method. We stored these behaviors in a robot’slong-term memory, and used a search mechanism togenerate autonomous control based on robot's perceived
Page 3: [16] in that each behavior is defin

Motion Generation for Humanoid Robots with ... - Brown University

Create successful ePaper yourself

Delete template?

Save as template?