Sharachchandra Bhat

I am a Research Assistant working with Dr. Sandeep Chinchali at the University of Texas at Austin. My research interests are in the application of Machine Learning for Robotics and Safe Autonomy using Formal Methods.

I received my B.Tech in Engineering Design and M.Tech in Automotive Engineering from IIT Madras. My Dual Degree Thesis was supervised by Dr. C S Shankar Ram.

Previously, I had worked at Systemantics India Pvt Ltd for three years, designing and producing industrial arms affordable to small and medium scale manufacturers in India. I was in charge of algorithm development for motion planning, controls and robot kinematics and dynamics.

Email  |  CV  |  LinkedIn  |  Github  |  Twitter

• May 2023: Graduated with a Master's from UT Austin.
• March 2023: "Co-design of communication and machine inference for cloud robotics" accepted to Autonomous Robots journal.
• August 2021: Joined UT Austin to pursue Master's in ECE with a major in Robotics

Large image compression gains can be obtained by focusing on downstream task performance at the cost of interpretability. Such a compression is suitable for machine perception but impossible to debug. Showed that the trade-off can be balanced by adversarial training an encoder-decoder network to give similarly high compression while providing human-interpretable reconstruction. Moreover privacy is preserved with reconstruction of images to a single class dependent code.

Robot teleoperation and cloud control is reliant on a strong and stable communication network. To utilize standard control paradigms designed for systems without delay it is necessary to monitor the network latency and compensate for its detrimental effects.

We developed a shield that provides probabilistic safety guarantees for a remotely controlled robot in the presence of stochastic network delays. The minimally invasive shield monitors the network latency and overrides control commands when necessary to ensure satisfiablity of LTL safety specifications.

Trained a neural network regression model to achieve faster point-cloud registration of rasterized 2-D Lidar scans with comparable accuracy to search-based methods. Implemented a full autonomous stack to run on an F1/10th car in a mapped environment: global navigation via Jump Point Search A*, localization via Particle Filters, obstacle avoidance via Path Scoring, and local navigation via Optimal Control.

Imitation learning in robot manipulation is known to perform better with policy networks that take in state sequences. Benchamrking against RNNs, we evaluated the effects of transformer design choices like cross-modal attention, featurizer networks, and input sequence size on task performance.

• Implemented a task space jerk limited trajectory generation algorithm for point to point motions.
• Optimized motion parameters with constraints of robot bandwidth and actuator saturation.
• Explored various path blending strategies for position and orientation and devised algorithms satisfying practical real-time computation limitations.
• Derived and implemented higher-order continuity quaternion splines that balance computational complexity with desired smoothness.

• Developed and tuned robust discrete-time joint level robot axes controllers in embedded C.
• Modelled and experimentally identified parameters of joint friction and dynamic model.
• Derived and implemented inverse dynamics analytical model of a hybrid robot architecture for feedforward control.
• Developed real-time algorithm for the control of flexible joints via MEMS triaxial accelerometer sensor measurements.

• Mathematically modelled a novel hybrid (serial plus parallel) architecture with six degrees of freedom industrial robot.
• Formulated a computationally efficient line search algorithm to solve the robot inverse kinematics problem. Discoverd an analytical forward kinematics solution.

• Identified possible robot singularities poses and devised singularity tracking and avoidance algorithms.
• Maximized workspace for a hybrid robot architecture to compete with standard industrial serial manipulator designs.

• Developed a comprehensive calibration model for a 4 degree of freedom planar parallel manipulator taking into account all possible inaccuracies arising from manufacturing and assembly.
• Experimentally obtained task-space calibration data from a FARO laser track.
• Implemented a stochastic gradient descent optimization algorithm to identify calibrated model parameters, improving robot accuracy.
• Sensitivity analysis to identify critical parameters and reduce model complexity achieving analytical solvability.