You have already rated this page, you can only rate it once! Your rating has been changed, thanks for rating!
In another example, when people talk to a robot about an emotionally difficult, personal event, they feel more positive when the robot makes subtle body gestures and short sentences suggesting that it cares about their experience. When I started this research, I expected that people would want robots to be as precise and predictable as possible and just fulfill human commands. My research shows that people prefer a robot that predicts what they want, and acts a little bit ahead of time, even when that means that the robot might make mistakes when it takes a wrong guess about what the humans will do.
We expect the machine to take initiative, and to make its own judgment about the situation, instead of just waiting for our commands. I believe that this is an inherent part of our expectation from artificial intelligence. It has been suggested that differential game theory GT can be used as a framework to describe various interactive behaviours between a robot and its human user Some recent papers have studied how finite games can deal with limited knowledge of the partner 17 , 18 , but it is unclear how the available techniques should be transferred to the continuous control of the interaction with a robot.
A method to model the interaction force of a contact robot with its user has been introduced in ref. However, the mechanical connection between the partners means that their individual behaviours will interfere and may lead to instability.
Human-Robot Interaction: Status and Challenges.
This adaptive GT controller is validated on reaching arm movements. It does not require measurement of the interaction force, and is proved to yield a stable interaction in typical conditions. The control can then be transferred to the joint space as described in the Methods. The variables used throughout this Article are listed in Table 1. Selecting the mechanical impedance process. In these equations, each player fulfils the reaching task by minimizing the error to the target while using minimal metabolic cost. Based on differential GT for linear systems 25 , the following control inputs of the robot and its modelled human user minimize the cost functionals equation 4 in the sense of the Nash equilibrium:.
However, it can be estimated from equation 3 without requiring force sensing that is, u h for actions along non-gravity-affected directions, while ensuring stability of the closed-loop system. Subtracting equation 3 from equation 6 then yields. For this purpose, we consider the Lyapunov function candidate. The above derivation leads to the following theorem demonstrated in the Methods that summarizes our results on co-adaptation and simultaneous control:. Competition can be implemented with a negative definite gain matrix Q for the robot in order to challenge the human user achieving their goal by pushing them away from it, which will promote active learning Different from collaboration, cooperation fixes the task performance through the total weight C , and uses equation 11 to share the effort.
As humans tend to relax during motor actions 27 , 28 , the robot will gradually take over the whole task load while the human will become passive. Another special case of cooperation particularly important in physical training is assist-less-than-needed 7 , 20 , designed to keep a human trainee engaged during sport practice or physical rehabilitation. This interaction control mode is implemented by setting C to make the robot short of reaching the target alone, which will induce the human to increase their effort in order to reach the target.
In the case of cooperation set also C. The robot dynamics were simulated using equations 5 , 6 and 10 to generate u , the simulated human dynamics used a similar set of equations to generate u h , and both of these inputs were used to drive the task dynamics according to equation 3.
Figure 1a shows that the collaborative GT controller yields a stable behaviour fulfilling the reaching task. The estimated control gains open circles converge in a few seconds to the real values solid line. Using two independent LQR for each agent dashed blue line , the target can be almost reached but with larger control gains. This leads to the simplified solution of two agents with independent LQR ignoring the interaction with the partner. Simulation of the reaching task, displayed in the dashed blue lines of Fig. We see in Fig. The GT controller considers the interaction with the simulated human through equation 5b and computes the minimal force for the robot to fulfil the task with them.
In turn, the simulated human can also minimize their effort. To test this question we carried out reaching simulations with a biased identification. Using these cost functionals with the values set in the Methods, both position and velocity gains are positive regardless of the estimation bias. However, the system matrix A is unstable due to damping.
We assumed that the simulated human improves their arm control by increasing the weight Q h trial after trial as described in the Methods. To promote active learning while still assisting a weak user, the simulated robot used an assist-less-than-needed cooperation strategy, with equation 11 and C making the robot short of reaching the target alone.
We see that while the reaching task is completed with similar performance in all trials Fig. Specifically, Fig. In a and b , the dashed lines correspond to real values and the solid lines to their estimation.
Inria Headquarters and research centres
Interestingly, it appears in Fig. This form of competition is desirable as it will let the user try to improve performance and be engaged in the physical training. However, the robot will not force the trainee beyond their motor capabilities, as it will keep compensating for the missing dynamics through equation In turn, this illustrates the capability of the adaptive GT controller to induce various types of interactive behaviour, namely here, less-than-needed assistance and competition.
The simulation results also illustrate that the two components in the weight matrices for reaching and viscosity that may correspond to spasticity in a neurologically impaired user 29 can be modulated separately, which makes it possible to implement requirements in various conditions. Considering the scenario of an impaired user such as a stroke survivor using the robot to train their motor control, their behaviour may be erratic with irregular progresses and setbacks over consecutive training sessions.
To test the capability of our interactive controller to deal with such erratic behaviours, we simulated a human user with performance varying trial after trial, which is implemented by random changes in the corresponding cost function. Figure 2d shows that even in this situation the robot can adapt to the changing performance by updating its own cost function weight in a suitable way trial after trial.
How will our designed interaction behaviour and adaptation work with real human users?
- Human–Robot Interaction.
- Human-Robot Interaction: Status and Challenges..
- Human-Robot Interaction: Safety, Standardization, and Benchmarking - CRC Press Book.
- The future of human-robot interaction?
- TIER 1 ARTICLE TYPES?
- Math Review For Standardized Tests (1st edition)!
- Personal Nutrition (7th Edition).
We tested this by implementing less-than-needed assistance control during arm reaching on the robotic interface shown in Fig. The red traces in Fig. The pink traces exhibit the adaptation taking place in the healthy subject scenario where the robot assists a representative user instructed to complete the reaching task as best as possible.
We observe that the reaching task is effectively carried out, while the robot successfully identifies the increasing contribution of the human user and lowers its assistance correspondingly.
Interestingly, the jumps in the weights at the intervals when the movement direction changes stimulate the convergence of the estimated values to the real values. When the subject is instructed to keep the arm passive, in a weak arm scenario, the robot observes this behaviour with a low estimated weight Q h and increases its motor weight Q correspondingly. The target and cursor are displayed on the monitor during movement.
The natural motor variability in different movements is not prevented by the controller, which may facilitate learning. To test whether the behaviour with our GT controller depends on subject-specific parameters, we repeated the same experiment with ten healthy subjects, who carried out five trials with three back-and-forth reaching movements, in each of the healthy and weak arm conditions.
Figure 3d shows that the controller does not prevent natural motion variability in consecutive trials, which may facilitate learning 20 , This demonstrates that our algorithm is able to clearly identify the different behaviours of each subject and adapt the GT control correspondingly. This is further illustrated in Fig.
The result shows that the controller is able to catch these behavioural changes in a suitable way. Altogether, these results demonstrate the efficiency of the simple regulation of assistance during cooperation of the human and robot through the sharing rule equation 11 with the GT controller of equations 5 , 6 and Contact robots that physically interact with humans are being increasingly used in industry and for physical training, but they lack a systematic methodology to produce versatile interactive behaviours.
Typical rehabilitation robots to help limb movement training 7 , or intelligent industrial systems to support heavy objects against gravity and facilitate their manipulation 30 , use a controller independent of the human user.
These two simple interaction behaviours, where the robot either ignores the human operator control in co-activity or blindly follows it in master—slave control have led to successful systems for specific tasks. The adaptive properties of the control and specific behaviours were further shown in a robotic implementation with human users mimicking physical training for motor recovery. In turn, the algorithm introduced here may be used to infer how human motor control corresponds to GT, and to determine which characteristic behaviours are adopted during various tasks.
Human–robot interaction - Wikipedia
The simulation and implementation results demonstrate that the new adaptive GT controller can flexibly implement different representative behaviours of physical interaction between two agents as proposed in refs. These include the collaboration or competition between partners without a hierarchy, the cooperation of complementary partners, as well as assist-less-than-needed assistance promoting engagement and active training.
Therefore, this work achieved the programme of ref. In contrast, the GT controller presented in ref. It can be used to regulate the interaction between any two autonomous agents, such as two robots, or to design human—robot interaction as was demonstrated in this Article. It may further serve as a tool for modelling bimanual control with brain interhemispherical communication in both healthy individuals and neurologically impaired for example, cerebral palsy subjects, as well as for modelling the control of physical interaction in a human pair.
The experiments reported here illustrate that the presented algorithm can also be directly used as a dynamic environment for training impaired reaching behaviour, for example after a stroke.
Interesting properties of the adaptive controller for training arm movements include the dynamic interaction not preventing natural motion variability in consecutive trials, as well as the stable and less-than-needed motion assistance, from nearly complete motion guidance to competition; both of these factors are critical for successful neurorehabilitation 36 , The experiment demonstrated the adaptability of the robot control, yielding a challenging but supportive training environment and inducing active inconspicuous adaptation.
The rigid body dynamics of many serial mechanisms can be represented through an equation of the form Substituting update law equation 10 into equation 9 yields. According to equation 5b , A r is bounded, thus so are P , L and u.