Seminar on Applications of Artificial Intelligence in Safe

Seminar on, Applications of Artificial Intelligence in Safe Human -Robot Interactions

Contents Introduction Human Modelling Prediction of the Human Trajectory Reactive Control Scheme Conclusion References 2

Introduction Integration of both Robot’s and Human’s workspace. A new sensory system for modelling, tracking & predicting human motions within a Robot workspace Obtain a superquadric-based model of human using SOM. Assess the danger of the robot operations. A new reactive control scheme. 3

Human modelling Sensory system for modelling and tracking human motion. Safety mat. Data processing using SOM. Obtain body orientation and location. Model of the human body. 4

Four steps in Human modelling Step 1: The safety mat consists of a number of pressureactivated nodes. Each node on the mat has fixed coordinates. Under the human body weight, a set of nodes F = {(xj, yj)|j = 1, . . . , n} are activated across the mat. Step 2: This set is then clustered into two subsets F 1 and F 2, corresponding to each foot using a SOM network. Step 3: Using these subsets, the human body orientation and its location are derived. Step 4: This information along with average human body dimensions is then used in order to obtain a model of the human. 5

Safety mat Detects obstacles. Constructed using 2 rubber sheets having parallel wires. Pressure activated nodes. Each node on the mat has fixed coordinates. …. 6

Data processing using SOM Activated node set F needs to be first divided(clustered) into two subsets, F 1 and F 2 corresponding to each foot. SOM network seems a suitable candidate for clustering the data representing human footprints. Input to the SOM network is (xj, yj) pairs and produce output as (f 1, f 2) 7

Data processing using SOM Correct clustering TYPE A Incorrect clustering TYPE B 8

Connectivity of the Laplacian matrics of the given sample set. Type A sample set have 2 zero eigen values in its Laplacian matrics… 9

To convert type B to type A, uncertain nodes are need to be removed. l 1 and l 2 corresponding to the outer borders and orientation of each soles. lavg represents inner border of the two soles. 10

Obtaining body orientation It can be obtained from 2 subsets F 1 & F 2 α- average of sole orientation Centre of the body These values used to obtain human model Lines l. L and l. R lines connecting the centers of the forefoot to the heel in each sole, respectively. 11

Model of the human body Unduloidlike structure Variable cross section at various heights. 12

13

Prediction of the human trajectory Motivated by ordinary human-human interaction. Observe the pattern of the motion and predict the motion using ANN. 14

15

Reactive control Formulation of the Danger Index. DI(κ, v) = f. D(κ)f. V(v) Impedence based Reactive Control Strategy. -Threshold value for DI. -Repulsive force -Virtual damping torque. Cumulative PDI method Fp = kp. PDI(n 0, dp)up 16

17

Conclusion Study on a sensory system and reactive control scheme. SOM network and superquadric functions are used for human modelling. Human motion predicted using ANN. Reactive control scheme is developed. 18

References Nima Najmaei and Mehrdad R. Kermani “Applications of Artificial Intelligence in Safe Human–Robot Interactions”, April 2011. A. De Santis, B. Siciliano, A. De Luca, and A. Bicchi, “An atlas of humrobot interaction, ” Mech. Mach. Theory, vol. 43, pp. 253– 270, 2008. R. Bischoff and V. Graefe, “Hermes—A versatile personal robotic assistant, ” Proc. IEEE, vol. 92, no. 11, pp. 1759– 1779, Nov. 2004. American Nat. Standard for Indus. Robots—Safety Requirement, RIA/ANSI R 15. 06— 1999, 1999. 19