Clean spills and crumbs in the table is still challenging, as it requires palnning while reasoning over uncertain latent dynamics via high-dimensional visual observations. The main contributions are:
- Describe the uncertain dynamics of dirty particles on the table using stochastic differential equation (SDE).
- Use visual observations of the table state to plan high-level wiping actions by using RL, entairely in simulation built above.
- Use whole-body trajectory optimization algorithm for navigation in the environment and table wiping.
Ideas: Only use RL to generate high-level actions, then use trajectory optimization to finish that subgoal. Actually that is not subgoal, just action.
We can extend this experiments by using “subgoal”.