Coupled filtering in visual and motor spaces for reaching and grasping
Résumé
Background: For human adults, reaching a target feels like an easy task, but nevertheless involves many degrees of freedom and complex trajectories in the joint space. To fully understand human movement generation, as well as to develop efficient algorithms for humanoid or dexterous manipulation robots (that share the same constraints), overcoming the limits and drawbacks of classical control approaches is required. Aims: Without resorting to inverse kinematics (that may be excessively time consuming or lead to rough approximations), we want to show that it is possible to control high dimensional systems by simulating and predicting the outcome of local actions (forward model only), as long as the problem complexity is broken down into both the visual and motor spaces. Method: We rely on a bio-plausible and probabilistic method for generating reaching movements in complex settings. Specifically, we apply (approximate) Bayesian filtering in the visual and motor spaces. Visual filtering permits to define an initial rough trajectory in the operational space (avoiding obstacles), which is then refined by motor filtering, allowing direct control in the joint space while respecting joint limits. Results: The method was validated in simulation on a set of scenarios, with one or several targets (e.g. fingers on an object for grasping) to be reached with one or several arms/hands. With strong spatial constraints (e.g. obstacles), it succeeds in finding a trajectory where inverse kinematics methods fail. Conclusions: Relying on the visual space seems promising to reduce the complexity of movement generation in the motor space.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |