Centre of Intelligent Robotics-IIITA

13/06/2023

We are happy to post the acceptance of our paper at Advances in Robotics 2023. While walking around the beautiful IIITA campus, it is often possible for visitors to know their position using just the images that their eyes see. We make robots have the same capability using cameras alone, instead of relying on a high-cost LiDAR sensor.

Visitors can register the images of a place in the morning and use the same for finding their position in the evening. Inspired by the same, we train a deep neural network that takes an image pair in different domains and calculates the degree of similarity. So, when our robot looks at an image, it can match it against registered images in a different domain by using a deep learning-based local search around the expected position (from the last iteration). The outputs are passed through an Extended Kalman Filter.

The visitors need to often look at interesting things in the left, right and forward to ascertain the position, and therefore our robot also has 3 cameras. The center congratulates the authors Utkarsh Kumar, Rahul Kala, G C Nandi. Research sponsored by IHFC.

23/02/2023

Happy to share the online version of the paper "Mission planning on preference-based expression trees using heuristics-assisted evolutionary computation" by Rahul Kala published at the Applied Soft Computing journal. Here we generalize a TSP solver to take a structured program (expression tree with AND, OR, and Sequence nodes) as input, instead of asking the robot to visit all sites. Research funded by Science and Engineering Research Board (SERB).

The mission planning problem has so far been solved by using temporal logic and classic planning approaches that have an exponential computational com…

22/02/2023

Phd scholar Rohit Yadav from centre of intelligent robotics IIITA presented his work as a part of annual PMRF review meeting at Indian Institute of Technology, Delhi. Discussed the challenges in SLAM under adverse weather conditions and how the proposed algorithms could be a good start for the fellow researchers working towards the same problem.
Rohit is sponsored by navAjna Technologies Private Limited and Science and Engineering Research Board (SERB) under Prime Minister Research Fellowship (PMRF) which is managed by FICCI.

16/02/2023

We are happy to state that Prof G C Nandi, the head of the Center and Project PI and Dr. Rahul Kala from IIIT Allahabad are visiting the University of Illinois Urbana-Champaign to interact with the group of Dr. Katie Driggs-Campbell as a part of the collaborative visit funded by IHFC and National Science Foundation (NSF), USA.

It is raining ideas with sunshine bright plans being baked in this beautiful campus, as the delegates spend quality time with the best in the field. The center is excited towards this collaboration as the two groups march together towards a common goal.

12/01/2023

We are happy to share the online availability of our article at the Applied Intelligence Journal. The paper aimed at making a low-cost localization solution for self-driving cars in extreme weather conditions.

A SLAM algorithm makes the vehicle continuously drift, especially in low-visibility conditions like evening. So, we primarily solve the problem as a visual place recognition problem, dividing the continuous route into discrete places. A deep-learning model takes a morning place image (reference) and a live evening image (query) to check if the query image belongs to the reference place. We match a sequence of reference place images with a sequence of live images to reduce errors.

A trick is that the vehicle cannot be the airport at 10:00 AM and at the railway station at 10:01 AM due to temporal cohesion. We restrict the place output to be within a neighbourhood of the place output at the previous time step.

The above logic can sometimes make one lost. Imagine driving a long highway where everything looks the same for kilometres. One is unsure of the place until a landmark is found, which may not be within a neighbourhood of a sequence of wrong place predictions. We reserve a subset of globally distinctive places as landmarks to come to rescue when lost.

Finally, the place recognition outputs are discrete, while the localization pose is continuous. Within a place we use the SLAM algorithm to compute the continuous pose.

The center congratulates the authors Rohit Yadav, Vishal Pani, Arpit Mishra, Naman Tiwari and Rahul Kala. Research supported by navAjna Technologies Private Limited

Simultaneous Localization and Mapping (SLAM) in extreme lighting conditions and weather conditions is a challenging problem due to a lack of features that actively cause the vehicle to drift. The localization is not possible on maps produced by the ideal continuous as features in different domains r...

07/01/2023

We are happy to share the online availability of our paper at Robotica. A robotic guide leads a group of visitors, while the robot should track the visitors for decision making. We observed several times the visitors would be out of sight and used social models to predict the position. The algorithm also predicts if the visitors are following the robot, left the group due to boredom or re-joined the group. The center congratulates the authors Dr. Vaibhav Malviya and Dr. Rahul Kala

Socialistic 3D tracking of humans from a mobile robot for a ‘human following robot’ behaviour

31/12/2022

29/12/2022

[SLAM concepts] A follow-up of the Applied Intelligence paper. Here’s a run of a vehicle at our campus with the vehicle continuously computing its position (the localization problem, blue) and making a map (the mapping problem, black)

Mapping: Look at the corner of the building. Using a stereo camera, compute its position in the camera coordinate frame and thus (assuming a known pose of the camera, blue) in the world coordinate frame. Add to the map (black). Image (input), Location (blue, input) → Map (black, output)

Localization: Look at the corner of the building in the image and locate the same in the map (black dot). Out of all corners (black), a sub-set is observed (red). As an optimization problem, compute the camera’s pose (blue) such that the projection of observed corners (red) in the image is the same as observed. Image (input), Map (black, input) → Location (blue, output)

Simultaneous Localization and Mapping (SLAM): Look at the corner of the building in the image. For several such corners observed in a sequence of frames as the vehicle moves, as a joint optimization compute the position of the corners in the world (map, black) and the sequence of poses of the camera (localization, blue), such that the projection of the corners (black) in the sequence of images for the camera poses (blue) is the same as observed. Image sequence (input) → Map (black, output), Sequence of Locations (blue, output).

28/12/2022

The Center of Intelligent Robotics is happy to welcome a new member, Rayeesa Mehmood, who has joined the center as a Senior Research Fellow at the project funded by the IHub Foundation of Cobotics. Today morning we introduced Rayeesa to the human and robot members of the center. It is a pleasure to welcome Rayeesa to our gang, to join us for some crazy ideas, activities and research.

27/12/2022

[SLAM paper] We are happy to report the acceptance of our paper entitled “Locality-Constrained Continuous Place Recognition for SLAM in Extreme Conditions” at the Applied Intelligence Journal. We solve a frequent problem with the courier delivery drivers in our campus where the courier delivery drivers tell us what they see, and we need to interpret their precise pose to guide them further (the localization problem).

We had the ambitious plan to localize a vehicle using a vision camera only in poor lightning conditions, while research typically makes use of expensive lidar sensors. We solved the problem as a Visual Place Recognition solver, dividing a continuous route into sequential places and training a deep learning network to match place images across domains (day and evening). Having experience of guiding several drivers, ensuring they see the same landmarks in the same order, we used heuristics to make the Visual Place Recognition algorithm give the place outputs in the same sequence as should be encountered in the route. If the algorithm still makes them lost at our beautiful campus, we also find the way out on seeing globally distinctive landmarks at our campus that are world-renowned.

The center congratulates the authors Rohit Yadav, Vishal Pani, Arpit Mishra, Naman Tiwari, and Rahul Kala. Research funded by navAjna Technologies Private Limited, Science and Engineering Research Board and FICCI under the Prime Minister Research Fellowship for Doctoral Research.

We’ll post some interesting insights in the coming days.