Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying
Tying a shoelace knot is commonly seen as a milestone for young children as they learn how to use their hands to execute complex motions. Humans are able to consistently tie a shoelace knot on a wide variety of shoes, even if they haven’t seen the shoe before. In comparison, shoe tying is a problem...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2022
|
| Online Access: | https://hdl.handle.net/1721.1/139115 |
| _version_ | 1826205298032902144 |
|---|---|
| author | Tan, Michelle |
| author2 | Tedrake, Russ |
| author_facet | Tedrake, Russ Tan, Michelle |
| author_sort | Tan, Michelle |
| collection | MIT |
| description | Tying a shoelace knot is commonly seen as a milestone for young children as they learn how to use their hands to execute complex motions. Humans are able to consistently tie a shoelace knot on a wide variety of shoes, even if they haven’t seen the shoe before. In comparison, shoe tying is a problem that the robotics community is still very far from solving, since it breaks many assumptions of existing algorithms in robotic manipulation. Some key difficulties of getting a robot to consistently tie any shoe include the complex dynamics, the deformable nature of the shoelaces, the long time horizon, the dexterity needed to manipulate flexible objects, and the large variation between different shoes.
In this thesis, we make progress towards shoe tying by making a robot that is able to improve on a given demonstration by making it more robust to initial conditions of a shoe in simulation. This is motivated by the fact that when humans learn to tie a shoe for the first time, they are carefully taught a procedure for making the knot, including how to hold the shoelaces and how to be able to tell if a shoelace knot is good or bad. The impressive part is that they can quickly adapt this procedure to any shoe. In this thesis, I discuss the following three contributions towards refining a demonstration to work for any shoe. The first contribution is the development of an open-sourced configurable shoe simulator environment that allowed us to tie shoes completely in simulation. The second contribution is a formulation and evaluation of direct policy search via CMA-ES with the goal of optimizing a given shoe tying policy for robustness. The third contribution is a formulation and analysis for learning the dynamics and cost on a latent state and an evaluation of the learned model for control. We found that CMA-ES and learning latent approximate information states were both successful techniques. Both were able to stabilize a demonstration for robustness on initial conditions of the shoelaces >95/100 of the time. |
| first_indexed | 2024-09-23T13:10:47Z |
| format | Thesis |
| id | mit-1721.1/139115 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T13:10:47Z |
| publishDate | 2022 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1391152022-01-15T03:28:35Z Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying Tan, Michelle Tedrake, Russ Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science Tying a shoelace knot is commonly seen as a milestone for young children as they learn how to use their hands to execute complex motions. Humans are able to consistently tie a shoelace knot on a wide variety of shoes, even if they haven’t seen the shoe before. In comparison, shoe tying is a problem that the robotics community is still very far from solving, since it breaks many assumptions of existing algorithms in robotic manipulation. Some key difficulties of getting a robot to consistently tie any shoe include the complex dynamics, the deformable nature of the shoelaces, the long time horizon, the dexterity needed to manipulate flexible objects, and the large variation between different shoes. In this thesis, we make progress towards shoe tying by making a robot that is able to improve on a given demonstration by making it more robust to initial conditions of a shoe in simulation. This is motivated by the fact that when humans learn to tie a shoe for the first time, they are carefully taught a procedure for making the knot, including how to hold the shoelaces and how to be able to tell if a shoelace knot is good or bad. The impressive part is that they can quickly adapt this procedure to any shoe. In this thesis, I discuss the following three contributions towards refining a demonstration to work for any shoe. The first contribution is the development of an open-sourced configurable shoe simulator environment that allowed us to tie shoes completely in simulation. The second contribution is a formulation and evaluation of direct policy search via CMA-ES with the goal of optimizing a given shoe tying policy for robustness. The third contribution is a formulation and analysis for learning the dynamics and cost on a latent state and an evaluation of the learned model for control. We found that CMA-ES and learning latent approximate information states were both successful techniques. Both were able to stabilize a demonstration for robustness on initial conditions of the shoelaces >95/100 of the time. M.Eng. 2022-01-14T14:50:52Z 2022-01-14T14:50:52Z 2021-06 2021-06-17T20:14:29.436Z Thesis https://hdl.handle.net/1721.1/139115 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Tan, Michelle Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| title | Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| title_full | Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| title_fullStr | Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| title_full_unstemmed | Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| title_short | Stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| title_sort | stabilizing demonstration trajectories of linear deformable objects for robotic shoe tying |
| url | https://hdl.handle.net/1721.1/139115 |
| work_keys_str_mv | AT tanmichelle stabilizingdemonstrationtrajectoriesoflineardeformableobjectsforroboticshoetying |