Physical manipulation planning with differentiable closed-loop manipulation primitives

Principal Investigators:

Marc Toussaint

Team Members

Danny Drieß

Physical manipulation planning in escape room scenarios

Research Unit 3, SCIoI Project 39

The target of this project are escape room scenarios where the solution can be inferred directly from an understanding of basic physical principles (intuitive physics), rather than from trial-and-error exploration. That is, we consider situations where the agent can – in principle – directly predict promising interactions sequences based on the perceived scene and physical understanding.
Our previous work on physical manipulation planning addresses such cases: Given a full scene description and a generic model of physical interactions, it can find optimal paths to a goal subject to physical correctness.
However, our previous work is not fit yet to become a demonstrator for SCIoI: The methods only predicts a possible path, not a robust controller to execute such a path with real-world robotics.
The objective of this project is to bring the generality and computational strength of the
existing planning framework to real-world execution, namely by extending it to compute a sequence of robust control modes rather than only a nominal reference path. The output of planning is now a fully reactive strategy that can directly be executed on the real system and contribute to the SCIoI escape room demonstrator.
With this work we aim to make substantial progress in transferring the strong generalization of our physical reasoning methods to robust real-world execution and demonstration in the SCIoI escape room scenario.

User registration

You don't have permission to register

Reset Password