2021 Fiscal Year Research-status Report
Development of Lifelong SLAM System for Service Robots using Deep Semantic Information
| Project/Area Number |
21K14115
|
|---|
| Research Institution | Tohoku University |
|---|
Principal Investigator |
Ravankar Ankit 東北大学, 工学研究科, 特任講師 (40778528)
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Keywords | Navigation / Mobile Robots / Deep Learning / Semantic Mapping / Scene Understanding / Path Planning / SLAM |
|---|
| Outline of Annual Research Achievements |
This study aims to realize a lifelong navigation system for mobile robots in dynamic environments by considering semantic information. Trials were conducted using several sensors in simulation and real environments. In the simulation, modeling of everyday household items and essential features using computer vision were studied and applied to the simulation environment for feature extraction and deep learning training. Several datasets in dynamic scenarios were collected and tested using sensor fusion algorithms. Sensor selection and assembly of actual robot hardware was completed. Deep learning training for detecting and extracting semantic information from the scene graph is currently being studied. Moreover, tests in real-environment have started. Several results were published in conferences and journals.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
This study aims to realize the long-term navigation of robots in challenging dynamic scenarios by considering deep semantic information. Despite the COVID pandemic that severely impacted and delayed the procurement of crucial robot components and primary sensors, we were able to test many of the algorithms inside the simulation environment. Furthermore, we conducted an evaluation study to determine appropriate sensors and actuators for the robot by analyzing several sensor parameters in the simulation. Based on obtained results, experiments on real-robot is planned for the next FY.
|
| Strategy for Future Research Activity |
For the FY2022, the plan is to develop the simulation model and real hardware concurrently and start experiments in real scenarios with actual robots. Additional sensors that were unavailable in the previous FY are planned to be procured and tested for data acquisition. To realize the long-term navigation for mobile robots, tests in several dynamic scenarios, including low-light conditions and durations (several days to weeks), are planned with continuous data acquisition. The later stage of this FY aims to train a robust deep learning framework to study dynamic changes in the environment by utilizing the hybrid metric-topological framework described in the original submitted proposal.
|
| Causes of Carryover |
For experiments in challenging dynamic scenarios, we will procure additional sensors attached to the robot hardware, such as a stereo camera, motion capture system, motion trackers, and 3D Lidars. Additionally, we will train the dataset on deep learning workstations (GPU). Miscellaneous expenses include reserving experiment rooms, travel funds for international conferences, open access publication fees, conference registrations, and other expenses such as consumables for mechanical and electrical components required for the robot hardware development and computing units.
|
Research Products
(30 results)
