Learning to Sequence Robot Behaviors for Visual Navigation 

  • Trained a hexapod how to switch between low-level control policies synthesized using a central pattern generator (CPG). The decision is taken based on raw visual feedback of a monocular camera.
  • The underlying agent is a Deep Q-Network that takes in raw camera pixels and outputs a low level policiy for the hexapod to follow for a certain time horizon.
  • Teammates: Puneet Singhal and Tanmay Shankar.


Quadrotor Ergodic Coverage

  • Developed  and applied a sampling-based ergodic coverage algorithm to a quadcopter coverage problem. A quadcopter has to ergodically cover a domain given a distribution of the regions of interests (in this video the region of interest is a bridge) in the domain. 
  • Using Microsoft's AirSim simulator.
  • Teammate: Steven Shan 


Hierarchical Deep Reinforcement Learning for Sequencing Behaviors

  • Taught a turtlebot in a hierarchical reinforcement learning framework how to sequence a number of visually learned policies in order to complete a navigation task .
  • The high level policy and the low level policies are dueling double DQNs.
  • Deep Learning Course Project, Fall 2017


Learning Vision-based Navigation

  • Trained a hexapod how to navigate in an unknown environment from raw camera pixels using a dueling deep Q-network implemented in Keras.
  • Used gym-gazebo package as an environment for interacting with the hexapod.

Deep Reinforcement Learning for LiDAR-based Navigation in Dynamic Environments

  • Trained a wheeled mobile robot how to navigate, in simulation, in an unknown dynamic environment while avoiding obstacles from raw LiDAR data using a dueling deep Q-Network.
  • Used Keras with Tensorflow in Python to train the agent, and used gym-gazebo package as a simulation environment.

Omnidirectional Jumping Monopod (GOAT leg)

  • Implemented virtual model controllers on an omnidirectional jumping monopod emulating the behavior of a virtual spring.
  • Wrote the high level controller code in C++.
  • The low level motor controllers are written in C#, and communication between various sensors is via ROS.
  • Team members: Daniel Piedrahita, Puneet Singhal, Bocheng Yu.

Decentralized Multi-agent Path Planning

  • Implemlemented a decentralized version of the M* algorithm that scales up to any number of robots, and requires only local communication between robots near each other, without the need for any sort of communication with a central computer.
  • Teammates: Puneet Singhal, Sha Yi, Yifan Ding, Zhening Yang


Autonomous Navigation using a Hexapod

  • Integrated a navigation system for a hexapod in Gazebo using ROS (hector mapping + smnav packages).
  • The hexapod performs 2D SLAM using a LiDAR. A global A* path planner generates the paths for the hexapod. A central pattern generator coordinates the motion of the hexapod
  • Teammate: Parva Patel.

Painting via Ergodic Coverage

  • Developed and applied a sampling-based ergodic coverage algorithm on a painting task.  A 7 DoF industrial robot is given a certain image. It interprets the image as an information distribution. The robot then applies the ergodic coverage algorithm to "cover" this distribution. This enables the robot to paint the picture in its own "artistic" fashion. 


Person Following using a Hexapod

  • Used Intel Euclid sensor to enable a hexapod robot to track a specific person.
  • I modified the Intel realsense person tracking package in C++ using opencv to allow the robot to track a specific person wearing a specific color.

3D Reconstruction for Realistic Simulated Camera Motion 

  • Implemented a virtual camera panning algorithm in C++ using opencv. 
  • Given two images of the same scene from different angles, we are able to reconstruct any frame between these two frames. We do a 3D-reconstruction of the scene, followed by a 3D-2D reprojection onto a moving virtual frame. 
  • Teammate: Brad Saund

Shape-Based Geometric Motion Planning for an Underactuated Highly-Articulated System

  • Developed controllers for an 8-link snake floating in space. Each controller reorients the snake along a desired direction.
  • Keywords: geometric mechanics, geometric control, shaped-based locomotion.

This work was done as part of The Robotics Institute Summer Scholar program RISS 2015.