Todor Davchev

I am a Research Scientist at DeepMind where I work on Robot Learning. I did my PhD at the University of Edinburgh, where I was part of the Robust Autonomy and Decisions (RAD) group. I also work closely with prof. S. Schaal and F. Meier and have completed two Google X AI residencies (project Intrinsic) and a research scientist internship with DeepMind.

Broadly, my interests lie in the intersection of robotics and machine learning. In my PhD I focused on improving the sample efficiency and robustness of learnt models through utilising inductive biases in the context of robotics applications. In the past, I have studied ways of building such biases through learning modular and re-usable across tasks world models [1, 2]; demonstrations [3,4]; transfer learning [5]; and self-supervised reinforcement learning [6]. I have particular interest in fast contact-rich skill acquisitions, such as peg, gear and plug insertions. In other work I have also studied ways to improve 2D trajectory generation in crowded and collaborative settings as well as extracting representations that are task specific and robust to perturbations. With this, I am generally excited about projects combining ML/RL/DL and robotics, including working on algorithms, simulations, actual robot experimentation, and the relevant software/hardware environments.

News Job I have accepted a Research Scientist role at DeepMind under Jon Scholz’ robotics team.
News Talk Presented at the MIT CSAIL department on Fast skill acquisition with goal conditioned RL, hosted by Pulkit Agrawal
News Research Our paper Residual Learning from Demonstration was accepted at RA:L and ICRA 2022!
News Research Our paper Hindsight Goal selection for Long-horizon Dexterous Manipulation was accepted at ICLR 2022!
News Research New paper on Learning Time-invariant Reward functions with Meta-learning
News Workshop ICRA 2021 live! Our workshop on Learning-To-Learn for Robotics aims to provide an informative overview of the existing challenges in L2L for Robotics. Consider submitting (deadline 15th May) Website.
News Workshop ICLR 2021 live! Our workshop on Learning-To-Learn brings together neuroscience and machine learning experts to push the boundaries of the field. Website.
News Research Our paper Model-Based Inverse Reinforcement Learning from Visual Demonstration was accepted at CoRL 2020!