Learning to See before Learning to Act: Visual Pre-training for Manipulation
Does having visual priors (e.g. the ability to detect objects) facilitate learning to perform vision-based manipulation (e.g. picking up objects)? We study this problem under the framework of transfer learning, where the model is first trained on a passive vision task (i.e., the data distribution do...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2021
|
| Online Access: | https://hdl.handle.net/1721.1/129384 |
| _version_ | 1826200683916820480 |
|---|---|
| author | Yen-Chen, Lin Isola, Phillip John |
| author2 | Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory |
| author_facet | Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Yen-Chen, Lin Isola, Phillip John |
| author_sort | Yen-Chen, Lin |
| collection | MIT |
| description | Does having visual priors (e.g. the ability to detect objects) facilitate learning to perform vision-based manipulation (e.g. picking up objects)? We study this problem under the framework of transfer learning, where the model is first trained on a passive vision task (i.e., the data distribution does not depend on the agent's decisions), then adapted to perform an active manipulation task (i.e., the data distribution does depend on the agent's decisions). We find that pre-training on vision tasks significantly improves generalization and sample efficiency for learning to manipulate objects. However, realizing these gains requires careful selection of which parts of the model to transfer. Our key insight is that outputs of standard vision models highly correlate with affordance maps commonly used in manipulation. Therefore, we explore directly transferring model parameters from vision networks to affordance prediction networks, and show that this can result in successful zero-shot adaptation, where a robot can pick up certain objects with zero robotic experience. With just a small amount of robotic experience, we can further fine-tune the affordance model to achieve better results. With just 10 minutes of suction experience or 1 hour of grasping experience, our method achieves ∼ 80% success rate at picking up novel objects. |
| first_indexed | 2024-09-23T11:40:15Z |
| format | Article |
| id | mit-1721.1/129384 |
| institution | Massachusetts Institute of Technology |
| language | English |
| last_indexed | 2024-09-23T11:40:15Z |
| publishDate | 2021 |
| publisher | IEEE |
| record_format | dspace |
| spelling | mit-1721.1/1293842022-10-01T05:11:35Z Learning to See before Learning to Act: Visual Pre-training for Manipulation Yen-Chen, Lin Isola, Phillip John Massachusetts Institute of Technology. Computer Science and Artificial Intelligence Laboratory Does having visual priors (e.g. the ability to detect objects) facilitate learning to perform vision-based manipulation (e.g. picking up objects)? We study this problem under the framework of transfer learning, where the model is first trained on a passive vision task (i.e., the data distribution does not depend on the agent's decisions), then adapted to perform an active manipulation task (i.e., the data distribution does depend on the agent's decisions). We find that pre-training on vision tasks significantly improves generalization and sample efficiency for learning to manipulate objects. However, realizing these gains requires careful selection of which parts of the model to transfer. Our key insight is that outputs of standard vision models highly correlate with affordance maps commonly used in manipulation. Therefore, we explore directly transferring model parameters from vision networks to affordance prediction networks, and show that this can result in successful zero-shot adaptation, where a robot can pick up certain objects with zero robotic experience. With just a small amount of robotic experience, we can further fine-tune the affordance model to achieve better results. With just 10 minutes of suction experience or 1 hour of grasping experience, our method achieves ∼ 80% success rate at picking up novel objects. 2021-01-12T18:26:08Z 2021-01-12T18:26:08Z 2020-05 2020-12-18T18:39:35Z Article http://purl.org/eprint/type/ConferencePaper 9781728173962 https://hdl.handle.net/1721.1/129384 Lin, Yen-Chen et al. “Learning to See before Learning to Act: Visual Pre-training for Manipulation.” Paper presented at the 2020 IEEE International Conference on Robotics and Automation (ICRA), Paris, France, 31 May-31 Aug. 2020, IEEE © 2020 The Author(s) en 10.1109/ICRA40945.2020.9197331 Proceedings - IEEE International Conference on Robotics and Automation Creative Commons Attribution-Noncommercial-Share Alike http://creativecommons.org/licenses/by-nc-sa/4.0/ application/pdf IEEE MIT web domain |
| spellingShingle | Yen-Chen, Lin Isola, Phillip John Learning to See before Learning to Act: Visual Pre-training for Manipulation |
| title | Learning to See before Learning to Act: Visual Pre-training for Manipulation |
| title_full | Learning to See before Learning to Act: Visual Pre-training for Manipulation |
| title_fullStr | Learning to See before Learning to Act: Visual Pre-training for Manipulation |
| title_full_unstemmed | Learning to See before Learning to Act: Visual Pre-training for Manipulation |
| title_short | Learning to See before Learning to Act: Visual Pre-training for Manipulation |
| title_sort | learning to see before learning to act visual pre training for manipulation |
| url | https://hdl.handle.net/1721.1/129384 |
| work_keys_str_mv | AT yenchenlin learningtoseebeforelearningtoactvisualpretrainingformanipulation AT isolaphillipjohn learningtoseebeforelearningtoactvisualpretrainingformanipulation |