1.2 Human-Robot Interaction
1.2.4 Interaction Ergonomics
Stress at work is a major preoccupation for our societies and this phenomenon is not bound to a particular field. [Danna 1999, Ganster 2013] proposes multidisciplinary reviews on work stress and well-being at work. A review based on cost of work- related stress is presented in [Hassard 2018]. The cost of this social phenomenon is hard to estimate according to the authors. For [Rosch 2001], the health costs related to job stress is about $300 billion a year for American companies alone.
The origins of job related stress are multiple. In this work we are interested by the stress caused by robots. While safety remains the main criterion to enable humans and robots to share the same workspace, others emerge such as the level of stress and discomfort the human can feel in the vicinity of the robot. In this context, a robot should not cause excessive stress and discomfort to the human for extended periods of time. Some works have been conducted to evaluate the impact of cooperation with cobots on the mental safety of human’s beings. [Butler 2001] explores the interactions between humans and mobile personal robots. This paper
(a) (b)
Figure 1.11: Figure illustrating the notion of safety grid presented in [Sisbot 2007]. The robot takes into account that a human feels more threatened when he is sitting rather than standing.
describes the level of comfort the robot causes human subjects according to a variety of behaviors. These behaviors are defined by the robot’s speed, distance, design and more. In [Arai 2010], experiments are conducted to study the factors causing stress on a human operator working with a robot in a production assembly system. These factors can be the distance from a swinging robot to an operator, speed at robot’s movement towards an operator and so on. Similarly [Fujita 2010] studies the mental strains exerted by a robot hand over motion on a human operator.
In order to guarantee a certain level of comfort and safety for humans working near the robot, we can act mainly on two aspects of the interaction. The first aspect is based on the robot’s behavior and social considerations. The second one is based on the robot’s motion characteristics. A similar classification is made in [Lasota 2017]. In the following we present some of these works treating these two aspects of the interaction.
1.2.4.1 Behavioral ergonomics
The understanding of human social behavior constitutes an important field of study for roboticians. They attempt to extract the implicit rules and codes that define humans interactions in order to anthropomorphize a class of robots referred as so- cial robots (Sec.1.2.2.2). Such an approach can be used for the robot to anticipate Humans actions as a Human will be more efficient and more satisfied of the inter- action if the robot can anticipate his actions [Hoffman 2007]. It can also be used for a robot to communicate its intents. A user will feel more comfortable knowing the goal of the robot early in its movements [Dragan 2013b, Dragan 2013a].
The study conducted in [Lasota 2015] shows that human-aware robots increase the level of safety and comfort of the participants, while increasing their perfor- mances. The authors demonstrate that maintaining physical safety by simply pre- venting collisions, as they are about to occur can lead to low levels of perceived
Figure 1.12: A depiction of several similar studies ([Abend 1982, Morasso 1981, Bizzi 1984, Flash 1985]). The formation of humans or monkeys arm motions un- constrained in the horizontal plane is investigated.
safety and comfort among humans.
[Sisbot 2007] presents a human aware motion planner to generate not only safe and effective paths, but also socially acceptable paths. The planner takes into ac- count humans by reasoning about their posture and field of vision. For example a human will feel more comfortable if the robot operate in its field of vision rather than in its back. Similarly a human will feels less threatened when standing in relation to when he is sitting (See Fig 1.11). An extension of this work is presented in [Sisbot 2010] that takes into account the task constraints as well as human kine- matics.
These work are usually integrated at a decision level. It is also possible to design more acceptable motions by looking at their properties.
1.2.4.2 Motion ergonomics
Human societal behavior is being studied and mimicked to provide robots with social skills. The same can be applied to human’s gesture and movement. In the first instance the resulting works are integrated at the highest layers of planning. The robot movement can be adapted to respect implicit social rules (Fig. 1.11).
But it makes sense only if the motion has been designed beforehand at a lower level to satisfy ergonomics properties.
Since we are used to velocity profiles and trajectories similar to our own by interacting with other individuals throughout our lives, it is reasonable to think we are going to feel more comfortable when interacting with a robot reproducing human-like movements. Thus, as for the behavioral ergonomics, the research of motions with good ergonomics properties will use the human model as a source of inspiration.
Among the works covering this field, the model described by [Hogan 1984] is very popular. Referred as minimum-jerk model, its intent is to be an organizing principle for a class of voluntary movements. It implements the result obtained in a previous study on primates where the objective was to find a general principle for motion control ([Hogan 1982]). At the same time similar studies (Fig. 1.12) were conducted on primates and humans [Abend 1982, Morasso 1981, Bizzi 1984]. It was found that humans generate roughly straight paths with single peaked bell shaped velocity profile for point-to-point motions. This generalization is mainly true if the movement is expressed in Cartesian space, or task space. It is less generalizable in joint space. According to the authors of [Hogan 1984], the previous observations suggest that the underlying goal behind these voluntary movements was to make the motion as smooth and graceful as possible. To verify this theory they used dynamic optimization with an objective function that minimizes the square of the jerk over the duration of the movement, since maximizing the smoothness implies minimizing the jerk. The model’s outputs were relatively close to the experimental results and it was assumed that the minimum-jerk model provides a good general description of voluntary arm movements. For [Hogan 1984], the main quality of this model is to be an organizing principle, hence it can be generalized to most voluntary motion and provides superior predictive capabilities. In [Flash 1985] this model is extended to cope with multi-axis. It must also be mentioned that despite the quality of these studies, there were conducted on little panels of subjects.
The minimum-jerk model was built around the hypothesis that the human be- havior could be derived from a single organizing principle, which provides a simple generalizable model. However with this assumption comes a trade-off that is a lack of flexibility and adaptability. Yet flexibility and adaptability are some of the most researched feature for future generations of robots (Sec. 1.2.2.1). The model was also simplified to assume symmetrical velocity and acceleration profiles. However the observed results showed that the acceleration phase of a point-to-point movement is often shorter than the deceleration phase [Abend 1982, Morasso 1981, Bizzi 1984]. Different works confirmed that velocity and acceleration curves are asymmetric for a large variety of motions [Beggs 1972, Ostry 1987, Nagasaki 1989], especially for skilled motions. [Flanagan 1990] demonstrates that human movements cannot be generalized by bell-shaped and symmetrical velocity profiles. Similar results are presented in [Shin 2015] showing that human-like movements cannot be reduced to bell-shaped velocity profiles, but depend on the characteristics and the purposes of given tasks. It is also known that the minimum-jerk model fail to achieve its ob-
jective of reproducing human-like motions for curved paths. [Huber 2009] proposes the "decoupled minimum-jerk" model in order to enhance the ergonomic properties of the original model.
To resume human-like motion tends to be smooth, thus with constrained jerk. It cannot be generalized and its characteristics depend on its purpose. Furthermore it is assumed that robots with human-like movements will be optimal for human-robot interactions. The majority of research works hence focus on mimicking humans. Yet the pertinence of robots with their own motion properties can still be investigated. In order to generate suitable motions for HRI a part of my work during this thesis was dedicated to the research of motions possessing satisfying ergonomic properties.