Implicit shape representations for analytically differentiable contact dynamics
| Autor: | Widmer, Daniel |
|---|---|
| Přispěvatelé: | Geilinger, Moritz, Coros, Stelian |
| Jazyk: | angličtina |
| Rok vydání: | 2021 |
| Předmět: | |
| Popis: | With the advances in robotics and increased research in deep learning, a variety of differentiable physics frameworks have been introduced in recent years. This thesis aims at extending an existing simulation framework with the ability to handle collisions with complex geometries while preserving the property of differentiability. The penalty-based contact model is extended to handle frictional contact for various collision types by describing complex shapes with implicit shape representations. This approach allows a general formulation of the friction model for arbitrary signed distance functions. Two different ways for implicit shape representations are presented: an efficient method based on sparse volume hierarchies and a state-of-the-art deep neural network architecture for learning shape representations. The results are demonstrated and studied in simulation experiments and applied to inverse problems. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|