profile_picture

Akshara Rai

Robotics Graduate Student

About me

Hello! I am a graduate student at Robotics Institute, CMU. I primarily work on walking robots, but sometimes also on manipulators. I am interested in learning controllers for robots, while trying to take few trails and staying safe.

Currently, I am looking at ways of incorporating domain knowledge into machine learning frameworks for robotics. In particular, I am interested in the usefulness of simulations when learning on hardware.

Projects

Learning to learn from simulations

Work with Rika Antanova and Chris Atkeson

We are interested in finding ways of transferring information between simulation and hardware. This can be through hand-designed features from simulation or through neural networks trained in simulation. Once we have the information we want from simulation, we use it to learn controllers on robots in very few trials. We are also investigating sample efficient ways of updating our simulation based on data seen on hardware.

2014-present
Using Simulation to Improve Sample-Efficiency of Bayesian Optimization for Bipedal Robots
Akshara Rai*, Rika Antanova*, Franziska Meier, Christopher G. Atkeson

Paper Code

Learning to learn: Using simulation to expedite hardware experiments (Thesis Proposal)
Akshara Rai

Document Presentation

Bayesian Optimization Using Domain Knowledge on the ATRIAS Biped
Akshara Rai*, Rika Antonova*, Seungmoon Song, William Martin, Hartmut Geyer, Christopher G. Atkeson

Paper Video

Deep Kernels for Optimizing Locomotion Controllers
Rika Antanova*, Akshara Rai*, Christopher G. Atkeson

Paper Video

Sample Efficient Optimization for Learning Controllers for Bipedal Locomotion
Rika Antanova*, Akshara Rai*, Christopher G. Atkeson

Paper Video

Learning feedback terms for Movement Plans

Work with Franziska Meier, Giovanni Sutanto and Stefan Schaal

Movement plans are often generated offline and followed blindly. We are working on developing feedback terms that can be added to previously developed movement plans. This feedback helps reject disturbances that arise during execution, for example due to obstacles, tracking errors, etc.

2013-2014
Learning Feedback Terms for Reactive Planning and Control
Akshara Rai*, Giovanni Sutanto*, Stefan Schaal, Franziska Meier

Paper Video

Learning coupling terms for obstacle avoidance.
Akshara Rai, Franziska Meier, Stefan Schaal

Paper

Reinforcement learning for Robots

Work with Aude Billard and Tomohiro Shibata

In the past, I have worked on reinforcement learning algorithms for robots. For example, as part of my Masters, I worked on using failed demonstrations to come up with successful strategies. As part of an undergraduate internship, I also worked on a clothing assistant robot.

2012-2013
Learning From Failed Demonstrations in Unreliable Systems.
Akshara Rai, Guillaume de Chambrier, Aude Billard

Paper

Clothing Assistance with Dual-Arm Robot Using Reinforcement Learning.
Tomoya Tamei, Takamitsu Matsubara, Akshara Rai, Tomohiro Shibata

Paper

Education

I am a joint PhD student at CMU and MPI Tuebingen. Before this, I did my Masters at EPFL, Switzerland and undergraduate at IIT Kanpur, India.

Robotics Institute, CMU

Ph.D. Student

I am currently working towards my Ph.D. at the Robotics Institute at Carnegie Mellon University.

2014-present

Max-Planck Institute, Tübingen

Ph.D. student

I am jointly appointed as a graduate student at the Max-Planck Institute in Tübingen, Germany.

2014-present

École Polytechnique Fédérale de Lausanne

Graduate Student

I worked on my Masters in Microengineering with a Robotics and automation focus at EPFL, Switzerland.

2012-2014

Indian Institute of Technology, Kanpur

Undergraduate Student

I did my undergraduate studies at IIT Kanpur, India on Electrical Engineering.

2008-2012

Technical Skills

MATLAB and Simulink

Python, C++

Tensorflow and PyTorch

Embedded Systems

I work on Bayesian Optimization for learning controllers for robots. This involves working in physics-based simulations and on actual robot hardware. Most of our programming is done in MATLAB and Simulink, but I have used Python, C++ and C along with neural network libraries like Tensorfow, Caffe and Pytorch. Ask me for details and code!