Venkatraman Narayanan

I am a final year Ph.D. student in the Robotics Institute at Carnegie Mellon University, advised by Dr. Maxim Likhachev. My general research interests span Artificial Intelligence (AI), Perception, and Planning for Robotic Systems, including personal assistance robots, autonomous cars, unmanned aerial vehicles, and manipulators in warehouse automation and flexible manufacturing.

My thesis work on Deliberative Perception introduces and adapts classical AI techniques for reliable 3D robot perception, with focus on object recognition and pose estimation. Earlier, I developed motion planning algorithms for autonomous manipulation of articulated objects, navigation in dynamic environments, and navigation under topological constraints.

Central to our work on perception and planning are heuristic search algorithms that we actively research. For instance, we have developed multi-heuristic graph search techniques that allow algorithms like A* to scale-up to complex and high-dimensional graph search problems encountered in robotics.

I interned at the Uber Advanced Technologies Center in Summer '15 and with the self-driving car team at Google X in Summer '13.

Before coming to CMU, I obtained my undergraduate degree in Electronics and Communication Engineering from College of Engineering, Guindy (CEG), Anna University, India.

Deliberative Perception

A recurrent and elementary machine perception task is to localize objects of interest in the physical world, be it objects on a warehouse shelf or cars on a road. In many real-world examples, this task entails localizing specific object instances with known 3D models. For example, a warehouse robot equipped with a depth sensor is required to recognize and localize objects in a shelf with known inventory, while a low-cost industrial robot might need to localize parts on an assembly line.

Most modern-day methods for the 3D multi-object localization task employ scene-to-model feature matching or regression/classification by learners trained on synthetic or real scenes. While these methods are typically fast in producing a result, they are often brittle, sensitive to occlusions, and depend on the right choice of features and/or training data. We introduce and advocate a deliberative approach, where the multi-object localization task is framed as an optimization over the space of hypothesized scenes. Our thesis is that deliberative reasoning--such as understanding inter-object occlusions--is essential to robust perception, and that the role of discriminative algorithms should mainly be to guide this process.

Multi-Heuristic Graph Search

Many problems encountered in robotics, be it in perception or planning, lend themselves naturally to graph search formulations. Traditional graph search algorithms such as A* require a single "admissible" heuristic to guarantee solution optimality. However, designing such heuristics by hand is often challenging and time-consuming. Consequently, it would be beneficial to use a suite of independent "weak" heuristics or modern learning techniques such as deep neural networks to learn heuristics to guide graph search. To this end, we have developed a family of algorithms titled "Multi-Heuristic A* for searching with multiple inadmissible heuristics (in conjunction with one admissible heuristic) without compromising solution quality guarantees. Finally, these are abstract graph search algorithms applicable in many domains--for instance, we have used these in the contexts of both Deliberative Perception and high-dimensional robot motion planning.

Manipulating Articulated Objects

Personal robots need to manipulate a variety of articulated mechanisms such as doors and drawers, as part of day-to-day tasks. These tasks are often specific, goal-driven, and permit very little bootstrap time for learning the articulation type. In this work, we address the problem of purposefully manipulating an articulated object, with uncertainty in the type of articulation. We make two contributions: first, an efficient planning algorithm that, given a set of candidate articulation models, is able to correctly identify the underlying model and simultaneously complete a task; and second, a representation for articulated objects called the Generalized Kinematic Graph (GK-Graph), that allows for modeling complex mechanisms whose articulation varies as a function of the state space.

Planning with Topological Constraints

For a UAV on a surveillance mission, what is the optimal path that would enable it to circumnavigate particular regions of interest? How can one find paths for a ground robot that satisfy some constraint with respect to obstacles in the environment? These problems can be formulated as planning with homology and homotopy constraints, or more generally, planning with topological constraints. We present a framework based on graph-search to solve these problems, by capturing topological information in a graph state variable.

Planning in Dynamic Environments

Path planning in dynamic environments is significantly more difficult than navigation in static spaces due to the increased dimensionality of the problem, as well as the importance of returning good paths under time constraints. In this work, we develop an anytime planner that produces an initial solution quickly, and improves the quality of the solution as time permits. Additionally, by using 'safe intervals' rather than time as a state dimension, the planner can operate in real-time scenarios.

Journal Articles and Conference Publications

Abstracts/Workshop Publications