Programming human-humanoid haptic joint actions from human dyad models

Human-Robot Haptic Joint Action

   Our work challenges Human-Humanoid Robot Haptic Joint Actions. Such actions are illustrated by the Fig 1.4. They are actions that a human and a robot do together with a sustained contact, such as dancing [65], handshaking [5, 70], object manipulation [74], walking handin-hand, hugging [12, 64], arm wrestling, surgery [66], or exoskeletons enhancing a human’s physical capacities [59]. The important characteristic of Human-Robot Haptic Joint Actions is that the human and the robot interact physically, hence it is more often termed physical Human-Robot Interaction or pHRI. During pHRI, the human and the robot exchange mechanical power and heat, exert forces on one another (Newton’s third law) by means of a common contact surface. They also share common velocities and temperatures. Some researches focus on thermal interaction and heat transfer [6, 25]. Others focus on tactile interaction [4], where importance is given to the repartition of forces over the contact surface as for hugging. On the contrary, our work focuses on haptic joint actions where the interaction can be described by the net force and torque exchanged through the surface and the mean velocity of the points of the surface. For instance, during an object manipulation, we are mostly interested in the object trajectory; therefore, only the net force exerted on the manipulator’s end effector is of interest. In contrast, during a hug, the net force applied by each hugger is zero and yields no information;thus the repartition of the forces over the contact surface is important. Handshaking is a good mixed example. Although all the implementations reviewed in [5] focus on the net force, the pressure exerted on the hand and the temperature are also important: no-one likes to shake a cold hand or when the handshake’s grip is too strong (used to show or communicate dominance) [61]. Therefore it is primordial when studying pHRI to well define in what form energy is exchanged and what are the physic quantities that are exchanged and shared. In this work, we focus only on transfers of mechanical power, the mean velocity of the points of the surface of contact and the net force applied on it. We do not consider the tactile aspect nor the thermal transfers.

Studying pHHI

   pHHI and pHRI are strongly linked. It is certainly possible to design pHRI without inspiration and/or study from pHHI. The most widespread example is the car. However, it requires learning and it does not allow piloting a plane. Deriving interaction models from pHHI and implementing them into pHRI endows the robot with a behavior the human can understand/identify to. Thus the human is able to interact more intuitively with the robot,with few or no learning. For instance, it is easy for two humans to carry an object together on a predefined trajectory. Each partner assumes the other to have similar capacities and behavior as himself, giving a knowledge of the other partner’s limits and an ability to guess his intentions [60]. A partner might answer the question “What would I do in his place?” (empathy) to predict the partner’s intentions [37, 51]. However, Jarassé et al. [33] propose a more a global framework for pHRI where the robot not necessarily acts as an helper. Inspired from game theory, the robot could endow various behaviors, such as co-activity, competition,collaboration, assistance and education. Furthermore, pHRI is a powerful tool to validate pHHI models. Because identifying the parameters of a complex pHHI model can be difficult, it may be easier to implement the model on a robot to check that similar measures are obtained, and if similar behaviors can be reproduced. In [57], Reed studies how two humans collaborate to move a crank; he then implements a control law for a robotized motor to perform the task with a human. Reed conducts a Turing-like test where a human subject performs the task with the robotized motor but is told he is collaborating with a human. Most subjects believed that they are indeed collaborating with a human, and not with a robot. However, Reed is not able to reproduce with the human-robot pair all the behaviors observed during the collaboration between two humans. Reed’s work example also shows that pHRI might help to correct/complete the pHHI model, which in turn will help to implement better pHRI. The benefits of pHRI are not limited to the study of collaboration between humans. Robots are also used to study how humans physically interact with the environment, where they are used to simulate the environment [34]. The same robot can then be used to validate algorithms that implement human-like behaviors in the same environment [72, 73].

Robots living among/with Humans, Acceptability

   Besides, robots are more and more present in humans’ daily life, as Aibo from Sony and Nao from Aldebaran (Figure 1.6), and thus they gain a social dimension. As people like to pet a dog, they would like to interact physically with robots. Humanoid robots able of efficient pHRI would enable many new applications in domotics or service robotics. Because physical interaction requires close contact, it involves risk [41] and might cause more apprehension or fear. One cause may be that physical interaction is more intimate than verbal or visual interaction, but also more dangerous. For instance, when a human tries to pet a cat he does not know, he first does it carefully to not scare the animal and not being clawed. In a second time, once the apprehension is overcome, the human might pet the cat more generously. In a similar way, a successful pHRI would be very reassuring and helps forging a bond of confidence, a closeness, between the human and the robot. Therefore, the robot would be less perceived as a machine tool, or a mere object, but as a permanently present supporting partner or companion. Strong pHRI skills would increase robots’ acceptability in the society. One must not forget that a main characteristic of robots is that they are machines. Machines that can hurt people, such as cars, planes or chainsaws. Industrial and domestic accidents with machines are common. Robert Williams was the first man killed by a robot in 1979. But robot are also machines that can be broken if wrongly used. Therefore, it is necessary to build robots that can safely live among humans.

Table des matières

List of Figures
List of Tables
Introduction
1 Physical Human Robot Interaction: State of the art 
1.1 Human-Robot Haptic Joint Action 
1.1.1 Purpose of pHRI
1.1.2 Humanoid Robots
1.2 Impedance Control for pHRI 
1.2.1 Stability during pHRI
1.3 Collaborative Transportation Task
1.3.1 Proactivity
1.3.2 Enhancement considering the object’s geometry
1.3.3 Obstacle Avoidance
1.4 Collaborative Manipulation and Locomotion 
1.4.1 Passive Approaches
1.4.2 Proactive Approach
1.5 Conclusion
2 Monitoring human haptic joint action: 3 cases study 
2.1 Ground Reaction Force Measurement Systems
2.1.1 Design of Foot-located Force Measurement System
2.1.2 Choice for the study
2.2 Experimental Setup 
2.2.1 Tasks and Scenarios
2.2.2 With or without Visual Perception
2.2.3 Data Acquisition System
2.2.4 Subjects
2.3 Observations 
2.3.1 Data Visualization with AMELIF
2.3.2 Hip Swing Compensation
2.3.3 Gait Synchronization
2.3.4 Phases Decomposition of the Motion
2.4 Conclusion 
3 Programming human-humanoid haptic joint actions from human dyad models 
3.1 Introduction
3.2 Impedance based model for haptic joint action 
3.2.1 Context
3.2.2 Requirements and desired behavior
3.2.3 Notations and Hypotheses
3.2.4 Proposed Impedance Model
3.2.5 Behavior in Collaborative Mode
3.2.6 Summary and Discussion
3.2.7 Limits of this model
3.3 1D Follower Trajectory Planner
3.3.1 Object motion decomposition into primitives
3.3.2 Reactive Generation of Primitives Sequences
3.3.3 Implementation on the HRP-2 Humanoid Robot
3.4 Proactive 3D-Trajectory Planner 
3.4.1 Motion Primitives
3.4.2 Robot local frame
3.4.3 Reactive Generation of Primitives Sequences
3.4.4 Turning
3.5 Switch to Leader Mode with a Joystick 
3.6 3D Experimentation with the HRP-2 Humanoid Robot 
3.6.1 Scenario
3.6.2 Results
3.7 Discussion and Concluding Remarks 
4 User Studies 
4.1 Passive dynamic behavior of the hands and decoupled gait 
4.1.1 Improvements on HRP-2
4.1.2 Implementation of the passive controller
4.2 Proactive versus Passive Comparative Study 
4.2.1 Setting
4.2.2 Measurements
4.2.3 Questionnaires
4.2.4 Results
4.2.5 Conclusion of the Comparative Study
4.3 Impact of Training on Performance with Proactive 
4.3.1 Setting
4.3.2 Results
4.3.3 Conclusion
4.4 Comparison with Human-Human Collaboration Data 
4.4.1 Data Sets
4.4.2 Indicators
4.4.3 Results
4.4.4 Conclusion
4.5 Conclusion
5 Further Potential Extensions 
5.1 Sound Deprivation User Study
5.1.1 Conditions and setting
5.1.2 Results
5.1.3 Conclusion
5.2 Self-Stabilization through pHRI 
5.2.1 Evaluation of Stability
5.2.2 Benchmark test Scenario
5.2.3 Benchmark test with and without Stabilizer
5.2.4 Legs Impedance Control
5.2.5 Stabilization through Interaction
5.2.6 Discussion
5.3 Direct Physical Interaction
5.3.1 Guiding the robot by the hand
5.3.2 Handshaking
5.3.3 Discussion
5.4 Extension with vision-based control 
5.4.1 Cube detection
5.4.2 Visual Servoing
5.4.3 Results
5.5 Conclusion 
Conclusion
A Two partners Impedance
A.1 Dynamic Equation
A.2 Equilibrium positions
A.3 Energy of the system
A.4 General Case
Bibliography

Télécharger le rapport complet