Probabilistic Robotics
In the age of Deep Learning
Link: https://tum-conf.zoom.us/j/66800229985
Passcode: 502285
 |
 |
 |
| Wolfram Burgard | Fabio Ramos | Teresa Vidal-Calleja |
| University of Technology Nuremberg (Germany) | NVIDIA and University of Sydney (USA and Australia) | University of Technology Sydney (Australia) |
| Talk: Probabilistic and Deep Learning Approaches for Mobile Robots and Automated Driving | Talk: Stein methods for parallelized Bayesian inference in perception, state estimation and control
Abstract: ↵Uncertainty estimation is critical in all levels of robotics systems, from perception to control and sequential decision making. Bayesian inference provides a principled framework for reasoning about uncertainty but the computational cost of computing posteriors can make it impractical for deployment in robots. Fortunately, the recent availability of inexpensive, energy-efficient parallel computing hardware and differentiable programming languages has opened the possibility for the development of Bayesian inference algorithms that leverage parallelism and differentiability of both likelihood functions and priors to estimate complex posteriors. In this talk I will describe a powerful nonparametric inference method that uses both differentiability and parallelism to provide nonparametric posterior approximations in a timely manner. Stein Variational Gradient Descent and its generalizations can be used to formulate Bayesian extensions of common methods in robotics such as ICP for perception, particle filters for state estimation, and model predictive control for decision making. I will show that Stein inference scales better with the dimensionality of the data and can be implemented efficiently on GPUs. Finally, I will discuss extensions of Stein methods for sim2real and the automatic adaptation of simulators to reflect real observations. |
Talk: Probabilistic Crowd Flows for Socially Aware Navigation |
 |
 |
 |
| Niko Sünderhauf | Felix Frank | Ingmar Posner |
| Queensland University of Technology (Australia) | Volkswagen (Germany) | University of Oxford (United Kingdom) |
| Talk: The importance of uncertainty for reliable perception and action | Talk: Latent variable models empower probabilistic optimal control | Talk: Robots in Latent Space - Learning Capable Robot Models with Deep Generative Models |
 |
 |
 |
| Yarin Gal | Balaji Lakshminarayanan | Rika Antonova |
| University of Oxford (United Kingdom) | Google Brain (USA) | Stanford University (USA) |
| Talk: Uncertainty in Deep Learning: Lessons Learned from Medical Imaging | Talk: Plex: Towards Reliability using Pretrained Large Model Extensions
Abstract: ↵A recent trend in artificial intelligence is the use of pretrained models for language and vision tasks, which have achieved extraordinary performance but also puzzling failures. Probing these models’ abilities in diverse ways is therefore critical to the field. I will talk about our recent work exploring the reliability of models, where we define a reliable model as one that not only achieves strong predictive performance but also performs well consistently over many decision-making tasks involving uncertainty (e.g., selective prediction, open set recognition, calibration under shift), robust generalization (e.g., accuracy and log-likelihood on in- and out-of-distribution datasets), and adaptation (e.g., active learning, few-shot uncertainty). Plex builds on our work on scalable building blocks for probabilistic deep learning such as Gaussian process last-layer and efficient variants of deep ensembles. We show that Plex improves the state-of-the-art across reliability tasks, and simplifies the traditional protocol as it improves the out-of-the-box performance and does not require designing scores or tuning the model for each task. Paper, Blog |
Talk: Versatile active learning via focused Bayesian optimization |
| Time | Type | Speaker |
|---|---|---|
| 09:00 - 09:15 | Introduction | Rudolph Triebel |
| 09:15 - 09:45 | Talk | Wolfram Burgard |
| 09:45 - 10:15 | Talk | Rika Antonova |
| 10:15 - 10:45 | Talk | Balaji Lakshminarayanan |
| 10:45 – 11:15 | Coffee Break | |
| 11:15 - 11:45 | Talk | Fabio Ramos |
| 11:45 - 12:15 | Talk | Teresa Vidal-Calleja |
| 12:15 - 13:00 | Lunch | |
| 13:00 – 13:30 | Talk | Yarin Gal |
| 13:30 – 14:00 | Talk | Niko Sünderhauf |
| 14:00 - 14:30 | Poster Interactions | All the participants |
| 14:30 – 14:45 | Coffee Break | |
| 14:45 - 15:00 | Paper Presentations | Kshitij Sirohi |
| 15:00 - 15:15 | Paper Presentations | Norman Marlier |
| 15:15 - 15:30 | Paper Presentations | Gertjan J. Burghouts |
| 15:30 - 16:00 | Talk | Felix Frank |
| 16:00 - 16:30 | Talk | Ingmar Posner |
| 16:30 - 17:00 | Panel Discussion | All the speakers |
The award winning authors will be contacted separately for the prizes in coming weeks.
Our robots are envisioned to perform complex tasks in real, noisy, imperfect, partially observable, hardly perceivable and unexpectedly behaving environments. Due to the probabilistic nature of the physical world, it is crucial to endow robotic systems with the ability to handle uncertainties during the operation. Probabilistic robotics, a vibrant field that has gained enormous popularity since its inception in the mid-1990s, provides a compelling paradigm to address this dilemma.
The evolution of the idea of exploiting probabilistic techniques in robotics is paralleled in other relevant domains such as classical artificial intelligence (AI), which, in the era of deep learning (DL), still plays an important role on robot systems. In contrast, empowered by the learning-based data-driven characteristic, DL as a representative approach of modern AI, is able to remarkably enhance the functionality of our robots with its impressive performance on different robotic relevant tasks. However, each coin has two sides. The uninterpretable nature of DL and its data-hungry learning fashion has unfortunately posed limitations in the applications of various probabilistic approaches. Though there has been a plethora of research advances in pursuing a consistent and effective extension of this concept in recent years, the challenges arising when deploying our robots in varying scenarios still remain unsolved. In consideration of this, it is enlightening to take a look at the probabilistic techniques at the age of pre-DL and peri-DL from a probabilistic robotics point of view.
In light of the above, this workshop aims at providing a forum to bring together robotic and machine learning researchers as well as industrial experts with experience on developing probabilistic methods. The discussions will be centered on past achievements, current requirements, urgent challenges and future research directions and promising applications.
 |
 |
 |
 |
 |
 |
| Janis Postels | Jianxiang Feng | Jongseok Lee | Matthias Humt | Yizhe Wu | Rudolph Triebel |
| ETH Zurich (Switzerland) | DLR (Germany) | DLR (Germany) | DLR (Germany) | University of Oxford (United Kingdom) | DLR and TU Munich (Germany) |
This workshop is supported by the IEEE RAS Technical Committees on Computer and Robot Vision, and Algorithms for Planning and Control of Robot Motion. The organizers acknowledge the sponsorship from Agile Robots AG.
Should you have any questions, please do not hesitate to contact the organizing committee.