Embedded Files
Playing a major role in our lab, researchers study models and representations of robotic movement for a deeper understanding of robotic behavior.
Research Highlights
Research Highlights
Image adapted from Rouse et al. 2024 in Applied Sciences, MDPI
"Stable Heteroclinic Channel-Based Movement Primitives: Tuning Trajectories Using Saddle Parameters"
As robots and connected networks grow more complex, stable heteroclinic channels (SHCs) offer a promising method for defining motion trajectories. In this study, modifying individual saddle parameters proves to be successful in adjusting trajectory fidelity without disrupting the overall path.
Figure Description From the Publication:
Figure 3. (a) Saddle point kernels, desired trajectory, and produced trajectory for a square trajectory plotted in the task space. (b) The canonical state waveforms of the weighted kernels. The degrees of freedom are the x and y directions in Cartesian space. The initialization waveform of kernel one (the green waveform at time = 0) contributes to the part of the produced trajectory that deviates from the square. For the remainder of this work, the initialization and steady state error will be ignored.
Image adapted from Rouse et al. 2021 in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
"Visualization of Stable Heteroclinic Channel-Based Movement Primitives"
This study introduces an adjusted movement primitive framework that replaces the stable equilibria of dynamic movement primitives (DMPs) with saddle points. Results show that stable heteroclinic channels (SHC) - based movement primitives (SMPs) perform similarly. In addition, this research proves that the SMP kernel weight follows a given path when plotted in the task space.
Figure Description from Publication:
Fig. 7. (a) Circular trajectory produced using 10 kernels. (b) Heart-shaped trajectory produced using 10 kernels. (c) Heart-shaped trajectory produced using 20 kernels. The top row shows trajectories produced by approximating the weights using (9). The bottom row shows trajectories and weights after batch learning & optimization. The sampling approach reduces cost by an order of magnitude. The SMP cost of (b) top is 1.6% of the DMP cost using random weights; a similar difference can be seen for (a) in Fig. 4. This was run using SMPs and resembles the more complex trajectory-following tasks completed by Ijspeert et al. in [5]. The weight locations after learning are similar to the sampled weights, but show some adjusted spacing at sharp turns in the trajectory. More kernels are placed at these sharp points to achieve better trajectory-following.
Page updated
Report abuse