Our work at CMU's Robotics Institute working on HRI with CoBots this summer.

Thursday, June 25, 2015

Expressive Motion Space Study

Research of Laban Efforts as applied to robotic motion so far involves applying Laban Effort qualities to a static position or a set path from one point to another. However, instead of just varying behaviors of motion along this path, Laban Efforts could be used to plan out the path shape itself, using the concepts of indirect and direct motions in the Space motion feature (Knight and Simmons, 2009). As part of this research into Space, path shape and the pose of robot at the end point of the path would be tested for their importance in conveying attitude.

For example, while a direct path shape would entail a straight line from point A to point B, an indirect path shape might take the form of a wide arc that moves out to the right of the direct path before curving back to point B. An indirect path shape could also entail a rectilinear version of the arc, which would presumably form a 3 or 4-sided polygon with the direct path. The possible difference between an rectilinear indirect approach and a curvilinear indirect approach with respect to the robot's perceived attitude is one aspect of Space that could be explored as a part of this study. The indirect/directness of the robot's end pose would, in accordance with the research we've been doing in the past few weeks, rely on its distance from its attentional goals and its orientation to them.

Path shape as it relates to the legibility and predictability of robot motion has been researched in the context of a robotic arm moving towards attentional goals, with the conclusion that the predictability of a robot's movement does not always match up with its legibility, since legibility is based on an estimation of the robot's function based on its movement while predictability is based on an estimation of the robot's movement based on its function (Dragan, Lee, & Srinivasa, 2013). Predictability often coincided with direct-type movements while legibility coincided with more indirect-type movements. While the study in question focused on intent-expressive motion rather than attitude-expressive motion, it raises the interesting idea that the predicted movements of robots are likely based on people's current ideas of what 'robotic' motion looks like, which may be linked to the perception of robots as unexpressive and purely functional. It then follows that if robots behave unpredictably yet legibly, more human attributes would be projected onto them, including emotional attitudes.

The attitudes we'd like to test here are specifically intentional attitudes towards two separate attentional goals: one being some inanimate object, and one being the human observer. The inclusion of the human is intended to build upon past studies of human-robot interaction involving robotic gaze and motion, and to contrast how attitudes towards an agent are perceived differently from attitudes towards an inanimate object. Possible survey questions for the subjects of the study could include "How enthusiastic was CoBot about [the attentional goal]?" and "How friendly was CoBot towards you?"

The position of point B (aka the end of the path towards the attentional goal) is also an important consideration, since the distance and orientation to the attentional goal affects how the robot's intentionality to that object is perceived, as we've been studying for the past few weeks. With the incorporation of a human observer as a second attentional object, the calculation of point B and the path to it could take the human's position into account as well, since the movement of robots within/around "safety zones" of a human's personal space has been shown to affect their perceptions of the robot's attitude towards them (Dautenhahn and Sisbot et al, 2006). This being said, for the purposes of the study, the human's position with respect to the robot and the goal may need to be controlled somewhat so that the movement of the participant taking the study does not add too much variance into the data.

The issue of robot gaze also falls under the general label of Space effort, since it relies on variation of the robot's orientation. Although it may be too much to try to incorporate orientation manipulation during CoBot's planned path, the role of orientation in attitude expression could be tested once CoBot has traversed the path and arrived at point B. At that location, having CoBot assume a specific orientation towards either the human or the attentional goal (or neither) would presumably imply different states of attitude to the subject. Based on the findings of previous gaze research stating that a direct gaze makes people more likely to pay attention and subsequently interact with robots, it is likely that a direct gaze would also make people more likely to assign humanlike attitudes to them (Bruce, Norbakhsh, & Simmons, 2002).

Study design:
A tour guide scenario would nicely present a situation in which CoBot could logically demonstrate expressive attitudes both towards some interesting object as part of a tour, and towards the human that it is guiding. The study would, between participants, vary the directness of:
  • The CoBot's path towards the object
  • Its ending position in relation to the object and the subject
  • Its ending orientation in relation to the object and the subject. 
While it is moving along the path and after it has stopped, CoBot can be giving a spoken explanation of the object as per a usual guided tour. Afterwards, the subject would be given a survey of CoBot's attitudes toward them and towards the object.

Feasibility of project:
Code has been put in place to extract the position of the observing human for use in our program and to send trajectory commands to CoBot from our program once a list of waypoints has been formed. The next step is to exactly calculate what waypoints should be used given the starting location of CoBot, the location of the attentional goal, the location of the human, and an indirect/direct specification for motion effort.

Summary:
The ending position of the robot, the path shape of its approach to that point, and its subsequent orientation will all be manipulated with the Laban Effort system (specifically Space) to imply expressive attitudes from the robot about attentional goals.


Expressive Motion Time Study

The Laban Effort system provides a way of quantifying expressive motion in acting which can then be translated into expressive robot motions. These motions can be created, tested, and calibrated with the goal of creating more natural and lifelike interactions. After examining the Laban feature of space, specifically a robot's stopping distance from and orientation to an attentional goal, I am interested in exploring the Laban Effort feature of time with respect to a robot's approach to attentional goals.

The CoBot is a non-anthropomorphic robot with limited degrees of freedom for movement. Therefore its expressive motions must rely on low-level movements rather than high-level, human-like ones. The Laban Effort feature of time will be of key importance in eliciting expressive motion without an anthropomorphic body because the trajectory of the whole robot can move in time patterns. Humans tend to attribute emotional value to even the most abstract, inanimate objects given their timing and patterns of motion (Heider & Simmel, 1944).

Manipulations of time that will be explored include speed, acceleration, and temporal gestures. For example, speed or acceleration to and from a goal may lead human observers to make inferences about how excited the robot is about the goal that it approaches or how focused it is on completing the task. Temporal gestures such as a hesitation or pause before reaching the goal may lead to inferences about how confidence the robot is about reaching its goal or how invested it is in successfully reaching its goal.

Sharma, Hildebrandt, Newman, Young, and Eskicioglu (2013) show that in the Laban Effort system, the "sudden" time feature for robot flight paths (in this case moving to the next goal at a higher speed) was associated with a significantly higher reported valence and arousal of the robot than the sustained" time feature. This associates faster speed with a perceived stronger positive attitude in the case of a flying robot. Perhaps in different contexts, robot speed to attentional goals will elicit different reactions from humans.

Other research shows that changing speeds over the course of a motion (i.e. acceleration) can influence how natural a robot's motion appears. Compared to a constant velocity and a pattern of increasing-constant-decreasing velocity, a velocity that increases to the midpoint of time, then decreases was found to give the impression of smartness and familiarity (Mizoguchi, Sato, Takagi, Nakao, & Hatamura, 1997). Different patterns of acceleration may enable robots to seem to care more/less or have different attitudes about their goals.

Temporal gestures are another way to give animate characteristics to robots. In a study that used animated robots to explore the effects of robot anticipation and reaction (using hesitations and varied speeds in response to goals) robots were judged more favorably when they showed forethought and reacted to their success/failure at completing a goal (Takayama, Dooley, & Ju, 2011). Even the motion of inanimate objects with one degree of freedom such as automatic doors can make people perceive their "attitudes" in different ways (Ju & Takayama, 2009). Temporal gestures that relate to goals or stimuli are important for attributing emotion to robots.


-----
I will have to decide if I want to tackle all three: speed, acceleration, and temporal gestures or focus on one/two specific aspect of time. Temporal gestures are most interesting, but I also have to consider what is reasonable to implement (it seems like once I figure out the speed parameters all of three things should be fairly equally challenging, I think, because each of them will just involve determining sequences of waypoints)


Possible study designs
Robot is approaching some goal that it is meant to be checking -- because we want the robot to approach the goal and pay attention to it, but it does not need to collect any physical object or communicate with with it. Possibly a similar scenario to the hallway monitor checking plants will be used. Survey questions about the approach will ask about the robot's attitude toward the goal. The following are possibilities for manipulation. Some combination of the following will probably be used for the final design:
  • Different patterns of acceleration toward goal:
    • No acceleration, constant acceleration, constant deceleration, acceleration then deceleration, etc.
    • Possibly incorporate moving to as well as away from goal
      • accelerate toward goal then decelerate away and vice versa
  • Temporal gestures:
    • Hesitation before starting to move toward goal - possibly begin to move then stop to orient left and right (quickly v. slowly) before orienting toward goal and continuing toward it
      • This can be manipulated as "with hesitation" v. "without hesitation" or "quick" v. "slow" orientation or a combination of these factors
    • Hesitation on arrival at goal - Similar to above but at the end of navigation rather than beginning

Wednesday, June 17, 2015

Possible Project Topics


  • Exploration of the 'time' Laban state. How does variance of speed (and/or possibly acceleration) affect human perception of the robot's attitude towards attentional goals?
  • Implementation and testing of 'tour guide' features, including variance of path, gaze, speed, and position with relation to points of interest and the guided person.
    • Needs 'human detection' code
  • Perceived attitude towards agents vs. objects. How do you give robot or an object agency? Possible overlap with above topic.
  • Exploration of interaction between CoBots, and possible features of approach or sound that can be perceived by humans as social acknowledgement.
  • Exploration of robot's position relative to an agent or object in varying sizes of open spaces. How is the robot's proximity to you interpreted differently when in a large room vs. a small room? Study questions could include questions about the robot's likeability, insistence, intent.
  • Joint attention as an element of space?/flow?. Paper below talks about joint attention with an anthropomorphic robot, but how do we convey joint attention with a CoBot?
    • http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6005230
    • Needs 'human detection' code
    • May need 'human orientation detection' code
  • Exploration of path shape as it relates to Laban states and the perception of the robot as more 'natural'. In other words, how can altering the path shape (including orientation along that path) from point A to point B affect how humans perceive it?
    • Needed: follow trajectory code

Monday, June 15, 2015

Study Protocol - Calibration

Participants will be asked to rate how directly the CoBot is viewing an attentional goal. The view is direct if the robot seems to be paying attention to the object.

Participants will rate directness on a 5-point scale. Zero will be the center and and indicate a neutral stance. -2 will indicate a very indirect stance while 2 will indicate a very direct stance.

Participants will stand one meter to the right of the robot as indicated by tape on the floor. They will remain next to the robot for the duration of the study, walking up and down the line as the robot moves.

The robot will view the attentional goal at different orientations and distances. The participant will be prompted to rate the directness of each stance. Once they have rated one stance, they should indicate to the experimenters that they are ready to move on.



Behind the scenes:
  1. Run the program: python runLabanFeatures.py 1 <x y>
  2. The robot will move to 1 meter farther than start x position. This is the location of the attentional goal.
  3. Choose 'o' for orientation. Robot will move to first position. Place attentional goal on the spot it just vacated.
  4. 'a' and 'd' control orientations. Press 'a' to continually move through them. Press 'd' to go back. When you get to orientation 9, press 'a' once more to return to orientation 1, then press 'w' to move to the next distance.
  5. Repeat until finished with all 9 orientations at all 3 distances.
  6. Press 'esc' and run the program again, using the same x-, y-coordinates.
  7. The robot will move to 1 meter farther than start x position. This is the location of the attentional goal.
  8. Choose 'p' for position. Robot will move to first position. Place attentional goal on the spot it just vacated.
  9. 'w' and 's' control distance. Press 'w' to cycle through distances. 'w' will automatically change the angle and set the CoBot to its start position when it reaches its last distance.
  10. Be sure to record participants' ratings of each stance.

Calibration

To decide upon the parameters for calculating the target position and angle of the robot, we performed some calibration tests designed to assess our own understanding of 'directness' towards an attentional goal. Orientation was tested at three different distances and distance at three different orientations, to control for the interaction of these factors. Each distinct pose (orientation and distance from the attentional goal) was rated on a 5-point negative-to-positive scale with -2 being most indirect and 2 being most direct.

Orientation:
By averaging all the orientation data across all conditions, we produced a set of data that we mapped to a Gaussian curve, using the x-axis intercepts to determine where the threshold between 'direct' and 'indirect' orientations lay (and subsequently the value of 2σ for the Gaussian). We found that point to be 3π/16. 
Thus, to choose an direct target angle based on some desired center angle μ, we can sample from a Gaussian distribution where σ is 3π/32.

Similarly, we determined that for the double Gaussian structure of the indirect orientation distribution, the absolute value of both μ's relative to the center angle should be set as the midpoint between 3π/16 and π/2 (we've chosen π/2 as the upper limit of this range as we generally wish the robot to be turned towards its attentional goal in at least some capacity). This midpoint is equal to the center angle +/- 11π/32.

Sigma was then calculated from the basis that the difference between 3π/16 and π/2 (in other words, the desired range of distribution) should be equal to 4σ, thus making σ = 5π/64.

The graphed results of our target angle taken from over 500 different samples can be seen below, for both indirect and direct distributions.
As is shown above, there is a section roughly between 0.6 and 0.8 radians away from the center angle on both sides that the robot is highly unlikely to pick for either an indirect orientation or a direct orientation. These points fall inside a 'neutral zone' which often leads to more ambiguous interpretations of the robot's directness to the attentional goal. It is preferable that these angles are less likely to be chosen in order to provide a clearly interpretable example of an indirect or direct orientation.

Distance:
Similarly to orientation, we averaged all our data across all conditions to produce a set of directness ratings for each distance measure. Graphing this data and creating a line of best fit, we found the y-intercept to be 3.43 meters, representing the threshold where ratings of directness crossed from negative to positive. We set this value as the minimum proximity for determining indirect positions and maximum proximity for determining direct positions.

Wednesday, June 10, 2015

Game Plan

Phase 1 - Calibration

Write a program that presents various angles of orientation and positions to a plant which people will rate on a scale (indirect to direct) so that we can then determine what constants should be used in angle/position calculations (i.e. sampling from Gaussian distribution) to represent indirect, direct, and neutral states.

Phase 2 - Assessing individual features of Laban states

We will combine orientation and position to create approaches, pairing direct/indirect/neutral orientations and positions to see if any factor is dominating. The runLabanPilot program will call goToWithLabanFeatures. Each participant will assess multiple single approaches by rating the directness of the approach on a scale from 1 to 5.

Phase 3 - Laban state experiment

Robot has been trained to check the health of plants around Gates building so that it can later inform the Gates Gardening Attendant of which plants need tending. There will be 4 (2?) plants in total, half will be palm trees and half will be cacti. The robot will approach the plants directly, indirectly, or neutrally. The approach will be the same for the same type of plant for each participant. Participants will watch the robot approach each plant and then fill out a survey about perceived attitudes of the robot. Items will include questions about which plants seem to need attention, which plants the robot likes, etc.

Friday, June 5, 2015

Human Motion Observation

In order to understand how to apply principals of Laban Movement Analysis to robotic movement, we observed human movement in various settings. Specifically we focused on the Effort factors:


Laban EffortFighting PolarityInducing Polarity
SpaceDirectIndirect
WeightStrongLight
TimeSuddenSustained
FlowBoundFree


People walking in hallways or outdoor walkways:
  • Often turn their heads to look at things around them, especially when moving slowly. (Indirect, Free)
  • Often fix gaze on a single point when moving quickly. (Direct, Bound)
  • When walking alone, usually keep gaze toward the ground, looking up at items of interest and then returning gaze to the ground. ?
  • More often make eye contact with people passing when turning a corner rather than passing in parallel.
  • Tend to avoid obstacles (other people on the path) very far ahead of time by adjusting their trajectory. (Bound, Sustained)
  • Walk more slowly when travelling in groups. (Sustained)
  • Slow down when nearing possible collisions. (Sustained, Bound)
  • Veer slightly left or right when someone is behind them to allow space for the other person to possibly pass. (Indirect)
  • Cut close as they turn corners if they know where they are going and take wide (Direct), rounded paths around a corner if they are not sure where they are going (Indirect).
  • Tend to walk in the middle of a hallway when no one else is around (Indirect?) and to the right side if there are others in the hall (Direct?).
  • In groups expand and contract to fill the space they are in. (Indirect)
Campus tour guides:
  • Stand aside from attractions. (Indirect)
  • Walk slowly. (Sustained)
  • Take wide turns. (Indirect, Sustained, Bound)
  • Made lots of hand gestures. (Light, Free)
  • Slowly circle around to a stop. (Strong, Sustained, Bound, Indirect)

Sampling from single and double Gaussians

In order to pick an orientation angle towards the attentional goal, Cobot samples from a distribution of possible angles around the calculated center angle.

For a direct approach, we sample from a normal Gaussian distribution.

Data collected from 'direct' run of Laban motion program, featuring 100 samples around center angle -1.75 radians:

Data collected from 'direct' run of Laban motion program, featuring 500 samples around center angle -0.6 radians:


For an indirect approach, we want offset angles to be picked with higher frequency than those closer to the center angle. To do this we implemented sampling from a double-peaked Gaussian distribution where the peaks of highest frequency are offset symmetrically from the center by a constant distance of 1/8 radians.

Data collected from 'indirect' run of Laban motion program, featuring 100 samples around center angle -1.75 radians:

Data collected from 'indirect' run of Laban motion program, featuring 500 samples around center angle 0.75 radians: