Data-efficient Co-Adaptation of Morphology and Behaviour with Deep Reinforcement Learning
A novel approach combining deep reinforcement learning and design optimization in such a way that it is directly applicable to robots in the real world. This approach caters especially to roboticist who have access to only a single design/robotic prototype at a time. This paper is currently under review for the Conference of Robot Learning 2019.
In Review CoRL 2019 Deep Reinforcement Learning Optimization Adaptation of Morphology Evolutionary Robotics
Improved Exploration through Latent Trajectory Optimization in Deep Deterministic Policy Gradient
A Baxter robot learned to insert a cylinder, dangling from a string, into a tube using only camera images. I investigated here if we can learn a policy faster when using explorative actions generated from a trajectory optimizer. The optimization algorithm used a deep Q-network as objective function.
IROS 2019 Deep Reinforcement Learning Optimization Model-Based Exploration
From the Lab to the Desert
A sea turtle inspired robot manufactured from a paper laminate was trained to find the optimal locomotion strategy on different types of granular media, for example poppy seeds and real desert sand.
RSS 2017 Living Machines 2017 Reinforcement Learning Desert Bio-inspired Fast Manufacturing
Extracting Bimanual Synergies with Reinforcement Learning
This project investigated the latent spaces discovered throughout the training process of Sparse Latent Space Policy Search. We were especially interested in bi-manual synergies for lifting tasks. The Baxter robot received here only a reward and no sensory information about the environment.
IROS 2017 Reinforcement Learning Latent Space Sparse Latent Space Policy Search
Sparse Latent Space Policy Search
A novel reinforcement learning algorithm based on Latent Space Policy Search. We integrate Group Factor Analysis and reinforcement learning into one probabilistic learning framework. By creating several groups of actions we are able to learn specific latent spaces which either capture intra-group or inter-group relations.
AAAI 2016 Reinforcement Learning Latent Space Group Factor Analysis Variational Inference
Latent Space Policy Search
A novel reinoforcement learning algorithm which extracts automatically a low dimensional action space for better and faster exploration. The main idea is to learn the (probabilistic) principal components of actions which achieve high reward and explore this latent space during training.
IROS 2014 Reinforcement Learning Latent Space PPCA Expectation Maximization