Currently co-organizing a new real robotics challenge to advance the state-of-the-art in robotic manipulation, a key skill required to deploy robots in the real world. Where we host multiple robotic platforms at the Max Planck Institute for Intelligent Systems. Participants will submit their code as they would for a cluster, and it will then be executed automatically on our platforms. This will allow teams to gather hundreds of hours of real robot data with minimal effort.
CausalWorld is an open-source simulation framework and benchmark for causal structure and transfer learning in a robotic manipulation environment (powered by bullet) where tasks range from rather simple to extremely hard. Tasks consist of constructing 3D shapes from a given set of blocks - inspired by how children learn to build complex structures.
A framework of different derivative-free optimizers which can be used in conjunction with a model predictive controller and a learned dynamics model to control an agent in a mujoco or a gym environment.
Proposed a pipeline that comprises one of the most recent approaches, namely the HighResolution Network combined with a lowweight baseline model for extracting the 3D skeleton of human subjects applied to the Human3.6M dataset; splitting the challenge to image-related and geometric-related tasks.
With the increasing complexity of robots and the environments they interact with, crafting manual locomotion controllers becomes a challenging task to achieve. As evident from the recent research in this field, reinforcement learning (RL) contin- ues to emerge as an attractive solution for learning locomotion policies in high- dimensional continuous domains.
Most current exploration algorithms are based on heuristics and sampling/evaluating many potential viewpoints. The goal of this work is to learn an ideal motion to maximize the known space based on a local map representation.
Considered learning online adaptation in a model-based reinforcement learning context where we train a dynamics model, implemented as a Graph Neural Network, in conjunction with using MPC to control a system where the controller adapts to changes in the environment or tasks.
Successfully implemented a path planner and a velocity profiler for duckiebots as part of duckietown, a robotics outreach and educational platform, while taking into account various sources of uncertainty.
Successfully implemented a monocular visual odometry (VO) pipeline with the most essential features: initialization of 3D landmarks, key point tracking between two frames, pose estimation using established 2D ↔ 3D correspondences, and triangulation of new landmarks.
Successfully developed a Telepresence System that is equipped with high definition cameras remotely connected to an Oculus Rift. The purpose of the Telepresence System is to allow the users to see as if they were physically in place of the system itself. The project focused on three main areas: distortion removal, panoramic frame stitching and high definition video streaming.