This repo has code for the paper Active Learning of Abstract Plan Feasibility. In it we perform experiments in a block stacking domain where the evaluation tasks include constructing the tallest tower, longest overhang, etc. using a robot manipulator. This repo includes:
- A simulation-based particle filter to estimate the center of mass of blocks.
- An active learning strategy to learn a model for abstract plan feasibility with a information-theoretic approach to improve data sampling efficiency.
- (IN PROGRESS) A learning approach to estimate task-specific block properties using a learned latent representation.
All robot experiments (simulation and real-world) are conducted using a Franka Emika Panda robot.
For a more detailed overview of the repository and its folders, refer to the overview README.
See the full installation steps in the installation README.
For information on our hardware configuration, refer to our hardware setup README.
For specific information on the blocks used for our experiments, refer to our blocks README. This README also describes how to draw initial "home" block poses on the table.
See the full execution guide in the execution README.
Finally, refer to our troubleshooting README, contact us, or raise an issue if you are facing any problems.