Todor Davchev

Todor Davchev

PhD student in Robot Learning, University of Edinburgh

Posts by Collection

portfolio

publications

An Empirical Evaluation of Adversarial Robustness under Transfer Learning.

Published in International Conference on Machine Learning (ICML) 2019, Understanding and Improving Generalization Workshop, 2019

This paper studies the effects of using robust optimisation in the context of adversarial attacks. This allows us to identify transfer learning strategies under which adversarial defences are successfully retained, in addition to revealing potential vulnerabilities.

Recommended citation: Davchev, T., Korres, T., Fotiadis, S., Antonopoulos, N. and Ramamoorthy, S., 2019. An empirical evaluation of adversarial robustness under transfer learning. International Conference on Machine Learning (ICML) 2019, Understanding and Improving Generalization Workshop. https://arxiv.org/pdf/1905.02675.pdf

Vid2Param: Modelling of Dynamics Parameters from Video.

Published in IEEE Robotics and Automation Letters, 2019

This work shows how models trained entirely in simulation, in an end-to-end manner can perform online system identification, and make probabilistic forward predictions of parameters of interest in the phyical world.

Recommended citation: Asenov, M., Burke, M., Angelov, D., Davchev, T., Subr, K. and Ramamoorthy, S., 2019. Vid2param: Modeling of dynamics parameters from video. IEEE Robotics and Automation Letters, 5(2), pp.414-421. http://homepages.inf.ed.ac.uk/ksubr/Files/Papers/ICRA20Vid2Param.pdf

Residual Learning from Demonstration.

Published in arXiv preprint arXiv:2008.07682, 2020

In this work we propose residual learning from demonstration (rLfD), a framework that combines dynamic movement primitives (DMP) that rely on behavioural cloning with a reinforcement learning (RL) based residual correction policy.

Recommended citation: Davchev, T., Luck, K.S., Burke, M., Meier, F., Schaal, S. and Ramamoorthy, S., 2020. Residual Learning from Demonstration. arXiv preprint arXiv:2008.07682. https://arxiv.org/pdf/2008.07682.pdf

Model-Based Inverse Reinforcement Learning from Visual Demonstrations.

Published in Conference on Robot Learning (CoRL) 2020, 2020

In this work we propose model based inverse learning from visual demonstrations, a framework capable of learning cost functions from visual demonstrations.

Recommended citation: Das, N., Bechtle, S., Davchev, T., Jayaraman, D., Rai, A. and Meier, F., 2020. Model-Based Inverse Reinforcement Learning from Visual Demonstrations. Conference on Robot Learning (CoRL) 2020. https://arxiv.org/pdf/2010.09034.pdf

Learning Time-Invariant Reward Functions through Model-Based Inverse Reinforcement Learning.

Published in Arxiv, 2020

In this work, we propose a formulation that allows us to 1) vary the length of execution by learning time-invariant costs, and 2) relax the temporal alignment requirements for learning from demonstration.

Recommended citation: Davchev, Todor, Sarah Bechtle, Subramanian Ramamoorthy, and Franziska Meier. "Learning Time-Invariant Reward Functions through Model-Based Inverse Reinforcement Learning." arXiv preprint arXiv:2107.03186 (2021). https://arxiv.org/pdf/2107.03186.pdf

Learning Structured Representations for Trajectory Prediction in Crowded Scenes.

Published in IEEE Robotics and Automation Letters, Special Issue on Long-term Human Motion Prediction, 2021

This paper is about learning modular methods that explictly allow for unsupervised adaptation of trajectory prediction models to unseen environments.

Recommended citation: Davchev, Todor, Michael Burke, and Subramanian Ramamoorthy. "Learning Structured Representations of Spatial and Interactive Dynamics for Trajectory Prediction in Crowded Scenes." IEEE Robotics and Automation Letters 6, no. 2 (2020): 707-714. https://ieeexplore.ieee.org/document/9309332?source=authoralert

Wish you were here: Hindsight Goal Selection for Long-horizon Dexterous Manipulation.

Published in Deep RL Workshop @ NeurIPS 2022, 2021

In this work, we extend hindsight relabelling mechanisms to guide exploration along task-specific distributions implied by a small set of successful demonstrations.

Recommended citation: Davchev, Todor, Oleg Sushkov, Jean-baptiste Regli, Stefan Schaal, Yusuf Aytar, Markus Wulfmeier, Jon Scholz. "Wish you were here: Hindsight Goal Selection for Long-horizon Dexterous Manipulation." International Conference on Learning Representations (ICLR) 2022. https://arxiv.org/pdf/2112.00597.pdf

talks

teaching

Machine Learning Practical

Postgraduate course, University of Edinburgh, School of Informatics, 2017

I was heavily involved with the MLP course. My role involved helping and assisting with the course content, demonstrating, tutoring and marking.

Algorithms, Datastructures and Learning (Inf2b)

Undergraduate Course, University of Edinburgh, School of Informatics, 2018

Tutored 15 Undergraduate students on key symbolic and numerical technicalities regarding algorithms, data structures, probability, algebra and machine learning.

Informatics Research Review and Practice

Postgraduate Course, University of Edinburgh, School of Informatics, 2018

Lectured and marked 12 people on best practices to conduct current research in Machine Learning.

Decision Making in Robots and Autonomous Agents

Postgraduate Course, University of Edinburgh, School of Informatics, 2019

Marked the course on their understanding of models and techniques for decision making (under uncertainty) in robots. The emphasis of the course is on understanding how we can endow robots with the capacity to autonomously make decisions about how to interact with a dynamic environment (including, sometimes, other agents in these environments).