Haresh Karnan

I am a PhD candidate at The University of Texas at Austin, specializing in Reinforcement Learning, Computer Vision and Artificial Intelligence for Robotics. I'm advised by Peter Stone, and affiliated with the Learning Agents Research Group (LARG) at UT Austin Computer Science department. I'm also a member of the UT Austin RoboCup@Home team.

I was previously an Applied Science Intern at Amazon Scout where I worked on Deep Learning and Computer Vision for their package delivery robot.

In my past life, I worked with Dr. Robert Skelton at Texas A&M University, College Station, on state estimation and control of Tensegrity robots.

Email  /  CV  /  Biography  /  Google Scholar  /  LinkedIn

Robots navigating off-road environments typically require additional human feedback to learn operator preferences for visually novel terrains. In this work, we introduce a framework to extrapolate operator preferences from known terrains by leveraging multi-modality.

Visual terrain awareness is crucial for autonomous off-road navigation. We propose a novel way to learn relevant terrain representations from unconstrained, multi-modal robot experience collected with any policy. We show how such learned representations enable aligning off-road navigation behaviors with human operator preferences.

We introduce a large-scale, first-person-view dataset of socially compliant robot navigation demonstrations. SCAND consists of 138 trajectories, 25 miles of socially compliant navigation demonstrations collected on 2 robots by 4 human demonstrators within the UT Austin campus.

We introduce VI-IKD, a visual-inertial IKD model that learns to anticipate the kinodynamic interactions of the vehicle with the terrain ahead, enabling accurate high-speed navigation.

In this work, we introduce Optim-FKD, an approach for high-speed navigation that uses a learned forward kinodynamics model (FKD) coupled with non-linear least squares optimization for multi-horizon control.

Imitating an expert's video-only demonstration can be challenging in the presence of egocentric viewpoint mismatch. In VOILA, we introduce a unique viewpoint-invariant reward function to effectively imitate demonstrations on a physically different agent.

SOTA approaches in Imitation from Video-only demonstrations exhibit poor sample efficiency when learning with access to proprioceptive features of the imitator. To tackle this, we introduce VGAIfO-SO, an IfO algorithm that improves sample efficiency through self-supervision.

Sim-to-real is the problem of learning a control policy in an inaccurate simulated world such that the learned policy when transferred to the real-world, performs well. In this article, we explore the proposed black-box Sim-to-Real algorithm GAT, and its extension SGAT.

In this work, we propose the GARAT algorithm, which treats the Sim-to-Real problem as an Imitation from Observation (IfO) problem and uses advances in the IfO literature to transfer a control policy from a source domain to a target domain, using Adversarial Imitation Learning.

Real world dynamics are often stochastic and robot simulators have an inaccurate approximation of real world dynamics. In this work, we propose a Sim-to-Real algorithm called SGAT and transfer a Humanoid walk from Simulation to Real world.

In this work, we propose a Sim-to-Real algorithm called RGAT to ground an inaccurate simulator with data from the real world, using Reinforcement Learning.

Tensegrity mechanisms are known for their minimal-mass and flexible properties. In this work, we propose using vision based sensing for shape control of such soft robotic manipulators.

Visual terrain awareness is crucial for autonomous off-road navigation. We propose a novel way to learn relevant terrain representations from unconstrained, multi-modal robot experience collected with any policy. We show how such learned representations enable aligning off-road navigation behaviors with human operator preferences.

Robots navigating off-road environments typically require additional human feedback to learn operator preferences for visually novel terrains. In this work, we introduce a framework to extrapolate operator preferences from known terrains by leveraging multi-modality.

We introduce a standardized criteria and a metrics framework for evaluating social robot navigation in human occupied spaces. Key principles include safety, comfort, and social competency. This approach fosters fair algorithm comparison across diverse robots, simulators, and datasets, accelerating progress akin to computer vision and traditional robot navigation.

In this work, we propose the GARAT algorithm, which treats the Sim-to-Real problem as an Imitation from Observation (IfO) problem and uses advances in the IfO literature to transfer a control policy from a source domain to a target domain, using Adversarial Imitation Learning.

Template Credits