= Menu

Inferring Geometric Constraints in Human Demonstrations for Robot Programming

Guru Subramani
PhD thesis from University of Wisconsin - Madison — aug 2019
Download the publication : GuruSubramaniDissertation.pdf [26.3Mo]  

Many physical actions, such as pulling out a drawer or wiping a table may be modeled as geometric constraints. Geometric constraints restrict the motion of objects and generate reaction forces and moments when objects interact with them. Prior knowledge of the constraint is vital for robots to manipulate these constrained objects robustly. This dissertation focuses on methods to infer geometric constraints in human demonstrations by using both pose and wrench measurements. Specifically, it describes how wrench measurements can improve overall constraint inference.

The developed constraint inference approach addresses three problems: model selection (what the constraint type is), model regression (what the model parameters are) and demonstration segmentation (where they occur). It fits arbitrary scleronomic constraint models (constraints that impose restrictions on position and orientation independent of time) to kinematic and wrench measurements. It performs model selection by comparing models using position, orientation, force, and moment error metrics.

The constraint inference approach was evaluated using data from human participants performing tasks consisting of constraints. Experimental results conclude that wrench measurements improve both regression and model selection when compared to competing kinematic approaches. Wrench measurements help in situations where kinematic information is insufficient to determine the constraint, allowing constraint inference in short human demonstrations, typical in real tasks.

The inferred constraint geometry translate to constraint frames that are useful for robot control. Constraint frames map to hybrid position-force control parameters that enable robust robot interaction with the constraint. A proof of concept teaching by demonstration system called Robust Replay validates the utility of the constraint inference approach. Robust Replay converts a single human demonstration consisting of constraints into a robot program which is robust to deviations in the known geometry of the physical constraint. The approach decomposes a task into subtasks, identifies geometric constraints within those subtasks, and maps each subtask to an appropriate hybrid control description.

Images and movies


BibTex references

  author       = "Subramani, Guru",
  title        = "Inferring Geometric Constraints in Human Demonstrations for Robot Programming",
  school       = "University of Wisconsin - Madison",
  month        = "aug",
  year         = "2019",
  note         = "Defended June, 2019; Submitted August 2019",
  type         = "Ph. D. Dissertation, Department of Mechanical Engineering",
  ee           = "https://search.proquest.com/openview/bb94fd2f1c6899d0ce5a3c6e96180dd1/1",
  url          = "http://graphics.cs.wisc.edu/Papers/2019/Sub19"

Other publications in the database