IITK-TCS @ ARC-2017
Welcome !!
Once again we are participating in the Amazon Robotics Challenge 2017 to be held at Nagoya, Japan as part of the RoboCup 2017 event. This page is created to share our experience and knowledge with others.
Participating Institutes
Team Members
The team comprises of researchers, engineers and developers from both academia and industry. A brief profile of our members is available here.
Updates:
December 01, 2016
Our submission for ARC 2017. We have decided to use UR5 robot this year which is much lighter and faster compared to Barrett Arm that we used last year.
January 16, 2017
The Rack:
This time, every team has to design their own rack. We have decided to have 6 bins, some of them may have a different dimension compared to others. However, each one of them will be bigger than the minimum size of 31x20x14 cm. Maximum depth of our bins will be around 25 cm, this will reduce the length of end-effector, thereby simplifying the gripper design. The total volume for design 1 is 93,296 cubic cm, while for design 2, it is 94,600 cu. cm. The second design offers slightly bigger bins compared to the first design. The dimension of the storage (L, B, H) are within the limits of 125 cm. This rack will be mounted on a table or an adjustable stand as required for the task.. The thickness of wall is about 3mm and is made up Acrylic sheets to reduce weight.
Back Row (LtoR): Mohit, Swagat, Rajesh, Komal, Ravi, Dr. Behera, Samrat and Ashish. Front row: Siddharth, Dr. Venkatesh and Anima.
Picture Gallery
We will be uploading pictures over time. Keep visiting.
Result of ARC 2017
July 30, 2017
January 20, 2017
Our first experiment this year. We started with a RG2 gripper as the end-effector for a UR10 robot. The motion is generated using Moveit motion planning. We still need to include Octomap for collision avoidance. We have developed algorithm for primitive shape detection and finding graspable regions for the detected object. The object recognition is still done using RCNN.
Automatic Data Generation
February 20, 2017
Deep learning networks require a large number of training examples which are created manually which is a slow and laborious process. We are trying to automate this process of generating annotated data directly from the objects.
We secured fifth position in Stow task, third position in Pick task and fourth position in the final combined pick and stow task.
Publications & Patents
Design and Development of an automated robotic pick and stow system for an e-commerce warehouse. Arxiv Link.
Methods for Improving performance of robotic pick & place system for ARC 2017. ICRA Workshop on Warehouse Picking Automation 2017. [Paper][Slide]
Improve Motion Planning and Collision Avoidance
May 20, 2017
We are using Moveit for our motion planning work. The way point trajectory is generated using RRT motion planning algorithm. The rack itself (shown in green) is treated as an obstacle. The grasp pose direction is obtained from our grasping algorithm which is used by motion planner to generate the waypoint trajectory. The neighboring objects to a target is treated as an obstacle using Octomap. However, the use of Octomap is considerably increasing the motion planning time.
Grasping Algorithm
March 01, 2017
Here, we compute the graspable affordance and grasp pose from a single view of RGBD image obtained from a Kinect Camera mounted on the Robot. The algorithms uses RCNN image recognition window as the starting point, Surfaces are created using region growing algorithm on Surface normals. Empirical rules are applied to identify shapes and then
SMACH-ROS Architecture
July 07, 2017
We have now completed the migration of our code to SMACH-ROS architecture. This architecture simplifies task organization making it easier to deal with complex decision making process. The working of the system with this architecture is shown the following video. It also demonstrates the working on-line motion planning algorithm and collision avoidance using octomap.
Hybrid Gripper Design Version 3.0
July 19, 2017
We have improved our gripper design to some extent. The new gripper is more compact and faster in operation. It is now possible to pick the mesh cup using outward movement of fingers. The suction end point has a range of about 135 degrees of rotation. It has a payload of about 2 Kg. As you can see, we can pick the dumbbell using both suction as well as two finger gripper. However, the whole system is slightly bulky which is making it difficult for us to reach the corner regions of the rack. Another positive aspect of this gripper is that it is quite modular. The additional parts could be removed if required.
New Gripper Design
April 13, 2017
We have designed a new gripper that combines the suction with a two-finger gripper. Its length is around 40 cm with a payload capacity of 2 Kg. Two linear actuators are used to provide retractable action for gripper and swivelling action for suction end-point. This is an initial video that demonstrates the working of this hybrid gripper.
We stowed 18 objects out of 20 objects from the tote in about 12 minutes and dropped one on the way. Our gripper could not pick up the last two objects. We were fifth among 16 qualifying teams in this event.
Day 2: Pick Task Video
July 29, 2017
Day 1: Stow Task Video
July 28, 2017
We could successfully pick 9 items from the rack out of 10 target items and placed them into right boxes. This task was completed in about 7 minutes. it did not attempt for the last item due to a bug that was left during testing.
Day 3: Final Event Stow Task Video
July 30, 2017
Unique Visitor Count:
Final Event Pick Task Video
We were one of the eight teams who made it to the finals. We stow 12 out of 16 objects correctly and one erroneously. We blacklisted "white towel" and "ribbons" which could choke our suction gripper.
We picked 6 out of 10 target objects. Three of them were not picked up as the errors in the previous stow phase got carried on to this phase. We finished fourth in the final round.