I am very active in organising workshops at robotics and computer vision conferences.
In this workshop we will discuss the importance of uncertainty in deep learning for robotic applications. In addition, the workshop will host the 2nd Probabilistic Object Detection challenge, a new challenge that evaluates the ability of visual object detectors to accurately quantify their spatial and semantic uncertainty.
The workshop will provide tutorial-style talks that cover the state-of-the-art of uncertainty quantification in deep learning, specifically Bayesian and non-Bayesian approaches, spanning perception, world-modeling, decision making, and actions. Invited expert speakers will discuss the importance of uncertainty in deep learning for robotic perception, but also action.
This workshop brought together the participants of the first Robotic Vision Challenge on Probabilistic Object Detection, a new competition targeting both the computer vision and robotics communities. The workshop focussed on outcomes of the competition in the morning session, featuring talks by the four best scoring teams. After that we welcomed our invited speakers in the afternoon. Organised with Feras Dayoub and other colleagues from the Australian Centre for Robotic Vision, and Anelia Angelova from Google AI.
Deep Learning for Semantic Visual Navigation (CVPR 2019)
We discussed ideas to advance visual navigation by combining recent developments in deep and reinforcement learning. A special focus was on approaches that incorporate more semantic information into navigation, and combine visual input with other modalities such as language. Organised with colleagues from Google AI (Alexander Toshav, Anelia Angelova) and Imperial College London (Ronald Clark, Andrew Davison).
Similar in scope and topic to the CVPR workshop, here we discussed new benchmarks, competitions, and performance metrics that address the specific challenges arising when deploying (deep) learning in robotics.
In this workshop we discussed crucial challenges arising when deploying deep learning methods in real-world robotic applications, and a set of future large scale robotic vision benchmarks to address the critical challenges for robotic perception that are not yet covered by existing computer vision and robotics benchmarks, such as performance in open-set conditions, incremental learning with low-shot techniques, Bayesian optimisation, active learning, and active vision.
This workshop (led by Feras Dayoub) discussed the challenges of autonomous robots that have to reliably operate for long periods of time while having to demonstrate a high level of robustness and fault tolerance.
Learning for Localization and Mapping (IROS 2017)
The goal of this workshop was to present and discuss developments in learning-based approaches for localization and mapping systems.
Invited speakers: Wolfram Burgard (University of Freiburg), Jana Kosecka (George Mason University), Stefan Leutenegger (Imperial College London), Simon Lynen (Google), Marc Pollefeys (Microsoft).
Organised with Cesar Cadena and Igor Gilitschenski (both ETH Zurich), John Leonard and Sudeep Pillai (both MIT), and Fabio Ramos (University of Sydney)
New Frontiers for Deep Learning in Robotics (RSS 2017)
A wide range of renowned experts discussed deep learning techniques at the frontier of research that are not yet widely adopted, discussed, or well-known in our community. We carefully selected research topics such as Bayesian deep learning, generative models, or deep reinforcement learning for planning and navigation that are of high relevance and potentially groundbreaking for robotic perception, learning, and control. The workshop introduces these techniques to the robotics audience, but also exposes participants from the machine learning community to real-world problems encountered by robotics researchers that apply deep learning in their research.
Invited speakers: Yann LeCun (Facebook, NYU), Yarin Gal (University of Cambridge), Josh Tenenbaum (MIT), David Cox (Harvard), Chelsea Finn (UC Berkeley), Piotr Mirowski (DeepMind), Aaron Courville (Université de Montréal).
Organised with the support of Jürgen Leitner, Michael Milford, Peter Corke (QUT, Brisbane), and Pieter Abbeel (UC Berkeley).
Deep Learning for Robotic Vision (CVPR 2017)
Recent advances in deep learning techniques have made impressive progress in many areas of computer vision, including classification, detection, and segmentation. While all of these areas are relevant to robotics applications, robotics also presents many unique challenges which require new approaches.
Robotic vision specific challenges include the need for real-time analysis, the need for accurate 3d understanding of scenes, and the difficulty of doing experiments at scale. There are also opportunities robotics brings to computer vision, for example the ability to control position and viewing direction of the camera, and to provide a data source for “grounded” learning of concepts, reducing the need for manual labeling.
Invited speakers: Jitendra Malik (UC Berkeley), Raquel Urtasun (U Toronto / Uber ATG), Dieter Fox (U Washington), Honglak Lee (Google Brain / U Michigan), Abhinav Gupta (CMU), Jianxiong Xiao (AutoX), Richard Newcombe (Facebook), Raia Hadsell (Google DeepMind), Ashutosh Saxena (Brain of Things).
Organised with support from Jürgen Leitner, Michael Milford, Ben Upcroft, Peter Corke (QUT, Brisbane), Pieter Abbeel (UC Berkeley), Wolfram Burgard (Uni Freiburg).
We analysed why deep learning has not yet had the huge impact in robotics it had in neighboring research disciplines, and especially in computer vision. The workshop will identify the limits and potentials of current deep learning techniques in robotics, and will propose directions for future research to overcome those limits and realize the promising potentials.
Invited speakers: John Leonard (MIT), Larry Jackel (North C Technologies), Dieter Fox (Washington University), Oliver Brock (TU Berlin), Pieter Abbeel (UC Berkeley), Walter Scheirer (University of Notre Dame), Raia Hadsell (Google DeepMind), Ashutosh Saxena (Cornell and Stanford University).
Co-organisers were Jürgen Leitner, Michael Milford, Ben Upcroft, Peter Corke (QUT, Brisbane), Pieter Abbeel (UC Berkeley), Wolfram Burgard (Uni Freiburg).
This half-day workshop, co-organised with Ben Upcroft, Michael Milford (QUT, Brisbane), and Peer Neubert (TU Chemnitz), focussed on concepts and ideas for robust vision‐based place recognition in severely changing environments as well as discussing the extent to which place recognition is useful, or even required for robots.
This half-day workshop at ICRA 2015 in Seattle built on the highly successful 2014 workshop of the same name at ICRA, and discussed novel concepts and ideas for robust vision-based place recognition in severely changing environments. Organised with Peter Corke and Michael Milford.
Around 130 people followed the invited talks and paper presentations in a large ballroom.
This workshop continued the discussion from the previous year at ICRA and addressed the computer vision community at CVPR. Organised with Peter Corke and Michael Milford (QUT, Brisbane), and Torsten Sattler (ETH Zürich).
Approximately 40 people came by for talks and poster presentations. It was great to interact with the authors and of course the invited speakers Josef Sivic and John Leonard at CVPR as well as David Cox and Chi Hay Tong at ICRA. Thanks everybody for contributing!
Organised with Peter Corke and Michael Milford (QUT, Brisbane). We discussed novel concepts and ideas for robust vision-based place recognition in severely changing environments. Such changes – induced e.g. by the time of day, weather or seasonal effects as well as human activity – are a ubiquitous challenge for all autonomous systems aiming at long-term operations in both indoor and outdoor settings.
We had 9 contributed papers, a tutorial given by Peter, and invited talks by Michael and Paul Newman.
This full-day workshop brought together researchers working on novel approaches for modelling and inference in factor graphs. The goal of the workshop was to discuss techniques that introduce a larger robustness and allow incorporating multi-modal Gaussian or non-Gaussian measurements. New concepts of how to infere multi-modal posteriors were also in the scope of the workshop, as well as novel applications beyond the ubiquitous pose graph SLAM. I organised this workshop with John Leonard (MIT CSAIL) and Edwin Olson (University of Michigan)