Register now After registration you will be able to apply for this opportunity online.
Multi-View Representation Learning for Robotics
This Master's Thesis Project involves research and development of novel Machine Learning techniques for Representation Learning of a scene from multiple view points, for the purpose of Robotic Manipulation.
Various deep learning models have been developed to integrate computer vision into control pipelines for robotic manipulation. However, they all fundamentally learn representations that are strongly dependent on a fixed camera pose, and cannot generalize to novel camera views. This prevents wide deployment and reuse of pre-trained models in this field.
This Master's Thesis Project aims to develop a solution to this problem, using Contrastive Learning approaches to learn representations of visual data that are agnostic to camera viewpoints, and can be used in Reinforcement Learning or Imitation Learning pipelines for Robotics.
Applicants with a strong machine learning background are encouraged to apply. Experience with reinforcement learning simulation environments, robotics and computer vision are preferred.
Please include your transcript of records and cv with your application.
References:
https://arxiv.org/abs/2203.11024
https://arxiv.org/abs/2008.05711
https://arxiv.org/abs/2203.01983
Various deep learning models have been developed to integrate computer vision into control pipelines for robotic manipulation. However, they all fundamentally learn representations that are strongly dependent on a fixed camera pose, and cannot generalize to novel camera views. This prevents wide deployment and reuse of pre-trained models in this field.
This Master's Thesis Project aims to develop a solution to this problem, using Contrastive Learning approaches to learn representations of visual data that are agnostic to camera viewpoints, and can be used in Reinforcement Learning or Imitation Learning pipelines for Robotics.
Applicants with a strong machine learning background are encouraged to apply. Experience with reinforcement learning simulation environments, robotics and computer vision are preferred.
Please include your transcript of records and cv with your application.
Develop and implement a novel Contrastive Learning-based approach for Visual Representation Learning in the context of Robotic Manipulation. The learned representation should allow camera viewpoint-agnostic learning of manipulation controllers. The approach should first be tested in simulation and eventually on a real robotic arm. We expect the thesis to result in a publication in a Machine Learning or Robotics Conference.
Develop and implement a novel Contrastive Learning-based approach for Visual Representation Learning in the context of Robotic Manipulation. The learned representation should allow camera viewpoint-agnostic learning of manipulation controllers. The approach should first be tested in simulation and eventually on a real robotic arm. We expect the thesis to result in a publication in a Machine Learning or Robotics Conference.
Elvis Nava (PhD Student at ETH AI Center and Soft Robotics Lab) - elvis.nava@ai.ethz.ch
Barnabas Gavin Cangan (PhD Student at Soft Robotics Lab) - gavin.cangan@srl.ethz.ch
Elvis Nava (PhD Student at ETH AI Center and Soft Robotics Lab) - elvis.nava@ai.ethz.ch Barnabas Gavin Cangan (PhD Student at Soft Robotics Lab) - gavin.cangan@srl.ethz.ch
