This repository contains data as well as parsing and preprocessing scripts for the data collected in the user performance study. We performed this user study in order to evaluate subject performance in daily tasks that require physical interactions with objects and manuevering tools in a dexterous way.
| Dependencies |
|---|
| my_matlab_rosbag |
| WarpingTB |
The task investigated in this experiment was an asymmetrical bimanual task: "fruit scooping". The experiment was initially performed with 52 subjects, out of which only 37 had valid sensor data.
For recording demonstrations we used the following setup (also shown in the images below):
-
a 7 degrees of freedom (DoF) Kuka LWR robotic arm, and a Barret hand (4 DoF), were used for holding the mellon to be scooped. The demonstrator could kinesthetically guide the arm with his left hand. See Fig. 1.
-
a data glove covered with Tekscan tactile sensors on the front and side of the phalanges (see Fig. 2) was used to record the motion of the active hand performing the scooping motion.
-
the tool contained a 6-axis force-torque sensor and its motion was tracked with infrared markers (see Fig. 3).
-
an Optitrack Vision System was used to track the motion of the tool, human wrist and external objects (i.e. the trashing bowl, see Fig. 4). External cameras recorded the subject's performance (see Fig. 5).
In this experiment we collected the following data :
- sensor data, consisting of:
- robot data
- end effector position and orientation
- forces acting on the robot's end effector
- human hand data, consisting of:
- position and orientation of the wrist
- finger joint angles
- tactile readings from the patches on the fornt and side of each phalange and palm
- tool data, consisting of:
- tool position ond orientation
- forces and torques acting on the tool
- external objects - full pose tracking
- robot data
- post experiment questionnaire data (annonymised)
- video data - recordings of each demonstration performed by each subject
- video rating data - performed by external raters who have not done the experiment (only done for subjects with valid sensor data)
- external measurements - scooped quantity
The scripts in this repository perform the following steps for data preparation:
- ROS bag parsing
- smoothing
- rescaling
- dynamic time warping
- data plotting and visualization
The initial ROS bag files containing the sensor dat 5CBA a are not provided due to their large size (total 39.6 GB). We provide:
- the data obtained after bag parsing - see rawdata_sensors_parsed
- the data obtained after preprocessing and DTW - see ppdata_DTW
Video recordings are not provided in order to protect the subjects' identity (total size 8.69 GB).
For the analysis of the data in this experiment and the generated statistics see the following repository: user_performance_estimation
This work is currently under review: Pais Ureche, A. L., and Billard, A. (2017) Automatic skill assessment in learning from demonstration. (Under Submission, 2017)
Funded in part by the European Union Seventh Framewowork Programme FP7/2007-2013 under grant agreement no 288533 ROBOHOW.COG, and by the European Commission Horizon 2020 Framework Programme, H2020-ICT-643950.