Learning to See before Learning to Act

Implementation Notes

• This repo contains the work carried out as part of the Master's course Cognitive_Robotics at University of Groningen
• This repo contains a variation of the idea presented in Learning to See before Learning to Act paper, more details can be found in References
• Firstly, the training is done on passive vision task to learn to detect objects. We chose segmentation and in particular object grasp affordance segmentation instead of foreground segmentation as mentioned in the paper
• The trained passive vision task model is then transferred to learn an active vision task which is grasping

Pretrained model used for passive vision task [segmentation]

• The vision task used for learning is object part grasp affordance as segmentation
• Densenet-121 is the pre-trained model used
• Our code is in src/passive_task_segmentation/
• The dataset used for training passive vision task is UMD Grasp affordance segmentation RGBD dataset

Active vision task [grasping]

• The model which was learnt in the passive vision task is transferred to initialize the learning of active vision task of grasping
• A major portion of the code for learning grasping task is borrowed from GGCNN repository
• The modified code is in src/active_task_grasping/
• The dataset used for training active vision task is Jacquard sample dataset
• The main page of Jacquard dataset main page

Instruction to run scripts for passive vision task [segmentation]

• The src/passive_task_segmentation/requirements.txt lists all requirements required for passive vision task
• To list all training options

python3 src/passive_task_segmentation/train.py --help

• To list all inference options

python3 src/passive_task_segmentation/infer.py --help

• To visualize data samples and evaluate the performance of the passive vision task, use the notebook in src/passive_task_segmentation/helper_notebooks/passive_task.ipynb

Instruction to run scripts for active vision task [grasping]

• The src/active_task_grasping/requirements.txt lists all requirements required for active vision task
• To list all training options

python3 src/active_task_grasping/train.py --help

Simulation experiments to evaluate grasping performance

• For evaluating the performance of the trained grasping model, we used simulated YCB objects in PyBullet framework
• The original repo is available here - https://github.com/SeyedHamidreza/cognitive_robotics_manipulation
• The modified repo used for evaluation is here - https://github.com/AbhishekRS4/cognitive_robotics_grasping_manipulation

Contact info of team members

• Abhishek Ramanathapura Satyanarayana - a.ramanathapura.satyanarayana@student.rug.nl
• Amit Bharti - a.bharti.1@student.rug.nl
• Isabelle Tilleman - i.w.m.tilleman@student.rug.nl
• Nikos Douros - n.douros@student.rug.nl

References

• Learning to See before Learning to Act: Visual Pre-training for Manipulation
• GGCNN
• UMD Grasp affordance segmentation RGBD dataset
• Jacquard dataset main page
• YCB dataset