Suhan Shetty

Suhan Shetty

PhD Student | École polytechnique fédérale de Lausanne ‐ EPFL

Posts by Collection

portfolio

publications

Trajectory Tracking and Control of Car-Like Robots

Published in Machines, Mechanism and Robotics, 2018

This paper deals with trajectory tracking control of a car-like robot. By exploiting the differential flatness property of the system based on the dynamics, a trajectory tracking controller using flatness-based control techniques is designed. A singularity in the system for the chosen control inputs, which does not allow direct application of feedback linearization control, is identified and this singularity is overcome by applying the dynamics-extension algorithm to obtain a dynamic feedback linearized controller. This controller results in asymptotic tracking convergence of the system’s trajectory to the reference trajectory. Through numerical simulations, the control system is shown to track prescribed trajectories satisfactorily even in the presence of parametric uncertainties.

Recommended citation: S Shetty, A Ghosal, Trajectory Tracking and Control of Car-Like Robots, 2019 Machines, Mechanism and Robotics pp 759-767 https://link.springer.com/chapter/10.1007/978-981-10-8597-0_65

Trajectory based deep policy search for quadrupedal walking

Published in RO-MAN, 2019

In this paper, we explore a specific form of deep reinforcement learning (D-RL) technique for quadrupedal walking—trajectory based policy search via deep policy networks. Existing approaches determine optimal policies for each time step, whereas we propose to determine an optimal policy for each walking step. We justify our approach based on the fact that animals including humans use “low” dimensional trajectories at the joint level to realize walking. We will construct these trajectories by using Bezier polynomials, with the coefficients being determined by a parameterized policy. In order to maintain smoothness of the trajectories during step transitions, hybrid invariance conditions are also applied. The action is computed at the beginning of every step, and a linear PD control law is applied to track at the individual joints. After each step, reward is computed, which is then used to update the new policy parameters for the next step. After learning an optimal policy, i.e., an optimal walking gait for each step, we then successfully play them in a custom built quadruped robot, Stoch 2, thereby validating our approach.

Recommended citation: S Kolathaya, A Joglekar, S Shetty, D Dholakiya, A Sagi, S Bhattacharya, A Singla, S Bhatnagar, A Ghosal, B Amrutur, Trajectory based deep policy search for quadrupedal walking, 28th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN) https://ieeexplore.ieee.org/document/8956369

Trajectory Prediction with Compressed 3D Environment Representation using Tensor Train Decomposition

Published in ICAR, 2021

Trajectory optimization for motion planning is a fundamental problem in robotics. Algorithms typically require good initialization in order to find optimal trajectories. To provide such initialization, many approaches rely on the concept of memory of motion, where a function approximator is trained on a database of robot trajectories to predict good initial trajectories for novel situations, and hence speeding up the subsequent trajectory optimization process. To be able to generalize well to new environment, an expressive environment descriptor is necessary. We propose to encode the environment by discretized signed distance functions (SDF) which are then compressed using a tensor train (TT) decomposition approach. In order to show the expressiveness of this low-rank TT-SDF representation, three function approximators are compared: a nearest neighbor predictor, an artificial neural network and a mixture density network. We demonstrate the proposed method with motion planning examples on two different systems (point-mass and quadcopter). Our experiments demonstrate that the TT-SDF encoding can provide meaningful environment descriptors in order to predict good motion trajectories for warm-starting an optimal control solver.

Recommended citation: Brudermüller, L., Lembono, T.S., Shetty, S. and Calinon, S. (2021). Trajectory Prediction with Compressed 3D Environment Representation using Tensor Train Decomposition. In Proc. IEEE Intl Conf. on Advanced Robotics (ICAR), pp. 633-639 https://ieeexplore.ieee.org/document/9659407

Ergodic Exploration using Tensor Train: Applications in Insertion Tasks

Published in IEEE T-RO, 2022

By generating control policies that create natural search behaviors, ergodic control provides a principled solution to address tasks that require exploration. Since a large class of ergodic control algorithms relies on spectral analysis, they suffer from the curse-of-dimensionality, both in storage and computation. This drawback has prohibited the application of ergodic control in robot manipulation since it often requires exploration in state space with more than 2 dimensions. Indeed, the original ergodic control formulation will typically not allow exploratory behaviors to be generated for a complete 6D end-effector pose. In this research work, we propose a solution for ergodic exploration in multidimensional spaces using low-rank tensor approximation techniques. We rely on tensor train decomposition, a recent approach from multilinear algebra for low-rank approximation and efficient computation of multidimensional arrays. The proposed solution is efficient both computationally and storage-wise, hence making it suitable for its online implementation in robotic systems. We leverage our algorithm for high-dimensional ergodic exploration to solve the peg-in-hole insertion problem. We model peg-in-hole insertion task as a target detection problem and use our ergodic controller for the exploration in the 6D state space of the robot end-effector. The approach is applied to a peg-in-hole insertion task using a 7-axis Franka Emika Panda robot, where ergodic exploration allows the task to be achieved without requiring the use of force/torque sensors but only using human demonstrations. The approach can handle uncertainties in the location of the hole and/or graspings of the peg that typically exists in insertion tasks.

Recommended citation: S Shetty, J Silvério, S Calinon, Ergodic Exploration using Tensor Train: Applications in Insertion Tasks, IEEE Transactions on Robotics (T-RO) https://sites.google.com/view/ergodic-exploration/

Tensor Train for Global Optimization Problems in Robotics

Published in arXiv preprint, 2023

The convergence of many numerical optimization techniques is highly sensitive to the initial guess provided to the solver. We propose an approach based on low-rank tensor approximation techniques to initialize the existing optimization solvers close to the optima. The approach uses only the definition of the cost function (no gradient information is required) and does not need access to any database of good solutions. This renders the method less susceptible to getting stuck in poor local optima as compared to gradient-based methods. Unlike existing approaches in robotics that solve each optimization problem corresponding to a given task from scratch, we parameterize the optimization problem with respect to the tasks. We first transform the cost function, which is a function of task parameters and optimization variables, into an unnormalized probability density function. Then for a given task, we generate samples from the conditional distribution with respect to the given task parameter and use them as initialization for the optimization solver. As conditioning and sampling from an arbitrary density function is challenging, we use Tensor Train decomposition to obtain a surrogate probability model from which we can efficiently obtain the conditional model and the samples. The proposed method can produce multiple solutions coming from different modes (when they exist) for a given task. We first evaluate the approach by applying it to various challenging benchmark functions for numerical optimization that are difficult to solve using gradient-based optimization solvers with a naive initialization, showing that the proposed method can produce samples close to the global optima and coming from multiple modes. We then demonstrate the generality of the framework and its relevance to robotics by applying the proposed method to inverse kinematics and motion planning problems with a 7-DoF manipulator, as commonly encountered in robot manipulation.

Recommended citation: S Shetty, T Lembono, T Loew, S Calinon, Tensor Train for Global Optimization Problems in Robotics, arXiv:2206.05077 https://sites.google.com/view/ttgo/home

talks

An introduction to speech-based technologies for Natural Language Processing applications

Published:

Since the last two decades, the amount of data generated and collected has grown exponentially, and especially through the rise of unstructured data such as images, videos or text. More recently, audio and speech data have gained a large interest, for example through voice assistants. Companies like Google, Facebook, Apple, and Amazon have shown an increasing interest in professionals with skills and tools for ‘understanding’ and ‘transforming’ the massive flow of speech data in relevant information. Some of the most important speech-based technologies are voice activity detection, speaker diarization and identification, and automatic speech recognition. These techologies are often used as an input to various NLP applications afterwards. This brief workshop will give you a set of basic tools for grasping the main aspects of speech-based technologies and how they can be implemented in real-life cases.

teaching

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.