Academia.eduAcademia.edu

Figure 1 – uploaded by Vittorio Lippi

See full PDFdownloadDownload figure
Figure 1: A Snapshot of a 3Ball cascade training virtual environment (Lippi, 2008)

Figure 1 A Snapshot of a 3Ball cascade training virtual environment (Lippi, 2008)

Related Figures (7)

Figure 2: Model of Perception, action, cognition loop, when interacting with a training Virtual Environment (Avizzano, 2009);
Figure 2: Model of Perception, action, cognition loop, when interacting with a training Virtual Environment (Avizzano, 2009);
Figure 3: A sine shape noisy distribution, with the equivalent 8 Gaussian distributions approximation. features that make them very attractive, they work even with a limited number of samples, and they provide automatically to zero the output when the input is far from the dataset used for training.
Figure 3: A sine shape noisy distribution, with the equivalent 8 Gaussian distributions approximation. features that make them very attractive, they work even with a limited number of samples, and they provide automatically to zero the output when the input is far from the dataset used for training.
Model predictive controller is a formul ation to solve control problems which is finding increasing success within the scientific community (Camacho, 2007). integrates optimal control towards s making a proper use of an internal model of tl tochas The approach tic control, by he system to be controlled. The objective of the system, at each time is to control the input signal to be provided time, such as the future system dynamics op distance from the reference motion. foras hort window of timize a metric Figure 4: The motion experiment. The hand position is captured via Polhemus sensor, and the ball trajectory simulated with realtime physics. The experiment is repeated about 30 times. 8 state variables are monitored (hand/ball X/Y positions and velocities) at a frequency of 250Hz.
Model predictive controller is a formul ation to solve control problems which is finding increasing success within the scientific community (Camacho, 2007). integrates optimal control towards s making a proper use of an internal model of tl tochas The approach tic control, by he system to be controlled. The objective of the system, at each time is to control the input signal to be provided time, such as the future system dynamics op distance from the reference motion. foras hort window of timize a metric Figure 4: The motion experiment. The hand position is captured via Polhemus sensor, and the ball trajectory simulated with realtime physics. The experiment is repeated about 30 times. 8 state variables are monitored (hand/ball X/Y positions and velocities) at a frequency of 250Hz.
Figure 5: task analysis of the 1* experiment design, Four phase are present (start, launch, min, grasp) We have in mind the idea of finding a proper tool which is model generative (from observation) to recreate simple human-environment coordination tasks. In particular we have in mind to apply experimental analyses to ball juggling experiments (such as 3balls cascade).
Figure 5: task analysis of the 1* experiment design, Four phase are present (start, launch, min, grasp) We have in mind the idea of finding a proper tool which is model generative (from observation) to recreate simple human-environment coordination tasks. In particular we have in mind to apply experimental analyses to ball juggling experiments (such as 3balls cascade).
Figure 6: he four phases of motion: start (blue), launch (cyan), min(green), grasp (yellow).From top to down: XY position, Xvel vs X position, Yvel vs. Xposition how it influence the subsequent trajectories. In particular this missing analysis leads to a lack of adaptation during real task generation. This is a sever issue when the generated motion is meant to control any sort of underlying process for instance in the case of juggling the athlete progressively adapts to small translation/rotation of the balls in the air to better stabilize the overall process).
Figure 6: he four phases of motion: start (blue), launch (cyan), min(green), grasp (yellow).From top to down: XY position, Xvel vs X position, Yvel vs. Xposition how it influence the subsequent trajectories. In particular this missing analysis leads to a lack of adaptation during real task generation. This is a sever issue when the generated motion is meant to control any sort of underlying process for instance in the case of juggling the athlete progressively adapts to small translation/rotation of the balls in the air to better stabilize the overall process).
Figure 7:marginal plot of %,erurn(x, y) achieved through a Robbins-Monro approximation. in the training procedure (such as in SEDS approach). To understand better the behavior we plotted in Figure 7 a model of the x velocities learned from data by using Robbins-Munro interpolation (1952). The method gives an optimal solution for Gaussian approximation, and even if it is computationally heavy, given the reduced data set, it has been applied to check the regression outcome. Here we only have 1 hyper-parameter left which is the variance on the sampled data points. The hyper-parameters applies as a function smoother, thus allowing averaging between closer points and improving the interpolation between points.
Figure 7:marginal plot of %,erurn(x, y) achieved through a Robbins-Monro approximation. in the training procedure (such as in SEDS approach). To understand better the behavior we plotted in Figure 7 a model of the x velocities learned from data by using Robbins-Munro interpolation (1952). The method gives an optimal solution for Gaussian approximation, and even if it is computationally heavy, given the reduced data set, it has been applied to check the regression outcome. Here we only have 1 hyper-parameter left which is the variance on the sampled data points. The hyper-parameters applies as a function smoother, thus allowing averaging between closer points and improving the interpolation between points.
Figure 8:corrD MP example (Calinon 2010), inflated with 32 primitives instead of 8.The attractors diverge from the reference trajectory and only their combined behavior allows motion to stay in the exemplary motion. Besides the typical over-fitting issues which can be seen in the generated trajectories and solved by appropriate choice of hyper-parameters or regularization weights, we also found these learning algorithm being affected by a sort of masked oversizing. How can we be sure that the number of Gaussian chosen in a SEDS approach is adequate for the complexity of the trajectory? Are there hidden effects if more Gaussian than required have being employed? In figure 8 we reworked a Calinon (2010) example to ask DMP overfitting a simple trajectory, with more primitives (32 instead of 8). We observe that the DMP attractors, instead of being more and more adherent to the reference geometry, diverge from that thus achieving convergence on the desired trajectory only by reciprocal cancellation of dynamic effects. Even if the overall motion remains stable, this side effect is unacceptable in real systems for some control reasons: the numerical stability is achieved by difference of large attracting forces and hence may be affected to severe noises in case of reduced numerical precision, sampled time systems and quantized resolutions. Secondly, the cancellation is also accompanied with increasing stiffness that, as well known in the haptic literature, has severe stability problem due to sampling simulation time, non- modeled dynamical effects, and delays in acquisition and/or actuation timing. Finally, if we consider an energy function of the overall system:
Figure 8:corrD MP example (Calinon 2010), inflated with 32 primitives instead of 8.The attractors diverge from the reference trajectory and only their combined behavior allows motion to stay in the exemplary motion. Besides the typical over-fitting issues which can be seen in the generated trajectories and solved by appropriate choice of hyper-parameters or regularization weights, we also found these learning algorithm being affected by a sort of masked oversizing. How can we be sure that the number of Gaussian chosen in a SEDS approach is adequate for the complexity of the trajectory? Are there hidden effects if more Gaussian than required have being employed? In figure 8 we reworked a Calinon (2010) example to ask DMP overfitting a simple trajectory, with more primitives (32 instead of 8). We observe that the DMP attractors, instead of being more and more adherent to the reference geometry, diverge from that thus achieving convergence on the desired trajectory only by reciprocal cancellation of dynamic effects. Even if the overall motion remains stable, this side effect is unacceptable in real systems for some control reasons: the numerical stability is achieved by difference of large attracting forces and hence may be affected to severe noises in case of reduced numerical precision, sampled time systems and quantized resolutions. Secondly, the cancellation is also accompanied with increasing stiffness that, as well known in the haptic literature, has severe stability problem due to sampling simulation time, non- modeled dynamical effects, and delays in acquisition and/or actuation timing. Finally, if we consider an energy function of the overall system:
Related topics:
Human-Robot InteractionHuman Behavior Modeling
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved