Embedded Files
Tactile Perception – Pose Calibration
The point cloud projections of the soft bubble sensors intruded the object it was grasping. The goal was to develop a calibration pipeline that can be executed anytime the sensors are mounted for a more accurate point cloud representation.
We parameterize the rotation of the point cloud using the axis - angle representation while the position is given in Cartesian coordinates (x,y,z). We define a push-pull loss function using contact patch estimation and signed distance fields for improving pose accuracy.
Lane Change Detection on Free Ways
Detecting lane changes on free-ways using only GPS coordinates by modeling relative changes in the GPS coordinates to differentiate lane changes from road curvature.
Motivation: Treat road curvatures as background features in a video with lane changes as a foreground feature that changes more frequently and lasts for a shorter duration.
WBCs - Water Body Cleaners
An unmanned surface vehicle for treating water bodies of physical and chemical waste. Formulated a constant time position vector calculation algorithm and navigation and path planning for the vehicle.
Adding autonomous capabilities to an inexpensive off-the shelf drone
We worked on reverse engineering communication between the drone and it’s controller. The objective was to learn the rigid body dynamics for a quadcopter with a broken propeller a highly ill-posed physics problem. We used April Tags detected from the drone's camera for motion capture and computed the position and orientation of the drone to build a state - action space.
Vision-based Multi-Agent System
We track and command multiple blind robots (with no sensors). The aim was to design a central vision based planner navigate safely with multiple agents for selection and sorting of packages in a warehouse setup. The blind robots were used to avoid cost overheads for maintaining multiple sophisticated robots with expensive sensor configurations.
Affordance based Zone Mapping for First Person Action Recognition
We worked towards separating foreground and background in egocentric videos for a given frame by taking attention maps obtained from the gaze data and then inpainting the region obtained from attention masks to learn affordances without occlusion from the foreground features like the subject or the human hands.
Page updated
Google Sites
Report abuse