The program of this workshop includes 8 talks, spotlight presentations for the selected papers and a trajectory prediction challenge. In order to account for the diverse time zones of the invited speakers, the workshop will be split into the morning and the evening session in the Central European Summer Time zone (CEST). The talks will be recorded and all materials will be available on this website.

You can attend the workshop via this link: https://epfl.zoom.us/j/66418269303

Morning session

Time CEST (PST) Speaker Topic
09:30 - 09:45 (00:30 - 00:45 AM) Organizers Welcome and Introduction
09:45 - 10:15 (00:45 - 01:15 AM) Sami Haddadin, TUM Safe Motion and Interaction in physical Human-Robot Interaction
Robots need to be able to safely and sensitively interact with humans and their environment. This long standing goal of robotics research is still considered a grand challenge. However, recently significant leaps forward in physical human-robot interaction were made, including human-safe motion planning based on injury biomechanics. In the latter, the principles of mechanics are applied to the analysis of human response and trauma in order to understand how injuries happen at the level of the bones, joints, organs, and tissues of the body. Quantifying the response at those levels leads one to define injury tolerances at which humans fail to recover. Increased knowledge about the specific underlying injury mechanisms, understanding of human pain and injury probability together with associated tolerances allow embedding those data-driven models into the motion and control paradigms of collaborative mobile robots and therefore achieve guaranteed safe motion planning at maximum performance.
10:15 - 10:45 (01:15 - 01:45 AM) Dana Kulic, Monash University Human motion prediction from demonstrations and interaction
Understanding human movement is an important topic in biomechanics, human-robot interaction and rehabilitation engineering. In this talk, I will first describe our recent work on estimating human objectives by observing human actions and preferences, including time-varying objectives and and from incomplete observations, and illustrate how estimates of human objectives can be used for behaviour prediction. Next, I will describe how models of human movement during interaction can be estimated in-the-loop and used to guide interaction policy.
10:45 - 11:00 (01:45 - 02:00 AM) Break
11:00 - 11:30 (02:00 - 02:30 AM) Lihui Wang, KTH Motion Prediction for Human-Robot Collaborative Assembly
Human-robot collaboration has attracted increasing attentions in recent years, both in academia and in industry. For example, in human-robot collaborative assembly, robots are often required to dynamically change their pre-planned trajectories to collaborate with human operators in a shared workspace. However, industrial robots used today are controlled by pre-generated rigid codes that cannot support effective human-robot collaboration. In response to this need, human motion prediction is crucial for both collision avoidance and proactive assistance to humans, in addition to multi-modal robot control. Deep learning is used for classification, recognition and context awareness identification. Within the context, this presentation provides an overview as well as technical treatment on motion prediction for human-robot collaboration. Remaining challenges and future directions will also be highlighted.
11:30 - 12:00 (02:30 - 03:00 AM) TrajNet++ trajectory prediction challenge
12:00 - 12:10 (03:00 - 03:10 AM) Organizers Concluding the first session

Evening session

Time CEST (PST) Speaker Topic
17:00 - 17:10 (08:00 - 08:10 AM) Organizers Introducing the second session
17:10 - 17:40 (08:10 - 08:40 AM) Maren Bennewitz, University of Bonn Anticipating Human Movements and Foresighted Robot Navigation Using Learned Human-Object Interactions
Service robots that help humans in their everyday life should avoid interferences with them and adapt their navigation behavior accordingly. This requires a robot to anticipate the user's movements and to infer where support will be needed. In this talk, I present an approach that predicts the user’s navigation goal based on the robot’s observations and prior knowledge about typical human transitions between objects. Our system then generates a navigation strategy that minimizes the robot’s arrival time at the navigation goal and, at the same time, complies with the user’s comfort during the movement.
17:40 - 18:10 (08:40 - 09:10 AM) Jonathan P. How, MIT Context-aware learning of human motion prediction for safe autonomous driving
Understanding pedestrian behavior is crucial for safe autonomous navigation on the roads and in crowded environments. However, pedestrian motion is highly stochastic and involves various forms of interaction with the environment and other users of the road/sidewalk. While learning from large datasets is a promising strategy for developing predictive models, current methods are often limited to batch (offline) settings, thus limiting the ability for the model to generalize to new environments and update as new data becomes available. To address this challenge, this talk presents a Similarity-based Incremental Learning Algorithm (SILA) for pedestrian motion prediction. The talk will also discuss a new sharing pipeline called SimFuse, which enables autonomous agents to update their models by communicating with other vehicles (V2V) or infrastructure (V2I). Finally, the talk will describe our recent work toward certifying learned models to provide safety and robustness guarantees. This talk will conclude by highlighting some remaining open challenges in the field.
18:10 - 18:20 (09:10 - 09:20 AM) Break
18:20 - 18:50 (09:20 - 09:50 AM) Elena Corina Grigore, Motional Motion Forecasting for Autonomous Driving Applications
Autonomous driving requires operating in dynamic, interactive, and uncertain environments. Urban environments, in particular, present significant challenges for self-driving vehicles, which need to share the road with other agents such as vehicles, bicylists, and pedestrians. To do so, understanding the road context and making effective predictions of the behavior of these actors are crucial. Therefore, trajectory prediction is a core component of safe and confident operation. Among other difficulties, trajectory prediction is challenging due to the inherent multi-modality of this problem, with useful predictions needing to represent multiple possibilities and their associated likelihoods. In this talk, we present a solution for tackling multi-modality in a way that avoids mode collapse issues, and ensures we predict physically realizable trajectories. We do so by framing the trajectory prediction problem as classification over a diverse set of trajectory sets, and contribute three types of trajectory sets that can be employed in this context. Lastly, we also present a thorough comparison between our method and existing relevant work, contrasting regression, classification, and hybrid methods for trajectory prediction in self-driving.
18:50 - 19:20 (09:50 - 10:20 AM) Benjamin Sapp, Waymo Long term prediction in complex interactive environments
Long-term human motion prediction is a critical component of scalable autonomous driving systems. There has been a multitude of core modeling improvements for this task in recent years, in large part fueled by popular public benchmarks. However, existing datasets, metrics, and output representations leave much to be desired in their ability to capture interactions between agents. In this talk, we go over recent advances in modeling, datasets, evaluation and output representations at Waymo, with a particular emphasis on modeling interactions.
19:20 - 19:30 (10:20 - 10:30 AM) Break
19:30 - 20:00 (10:30 - 11:00 AM) Nick Rhinehart, UC Berkeley Towards Learning to Forecast Everything for Making Complex Decisions
I will describe our recent efforts on enabling learning-based forecasting methods to help agents make complex decisions. First, I will describe our method for tractable contingency planning by learning predictive behavioral models. Then, I will describe a method that offers a compelling alternative to the standard motion forecasting pipeline by inverting it: the first step in this method is to "forecast everything" (in our case, LiDAR videos). I will conclude with thoughts on scaling these methods towards a more general system for making complex decisions by learning to forecast everything.
20:00 - 20:40 (11:00 - 11:40 AM) Paper spotlight presentations
20:40 - 20:50 (11:40 - 11:50 AM) Organizers Concluding the second session

Accepted papers

We invite the authors of the accepted papers to present their work during the workshop. Each paper will get a 10 minute slot, which will include a 7 minute talk and 3 minute Q&A session, to present for the audience of the workshop.

Time CEST (PST) Authors Paper
20:00 - 20:10 (11:00 - 11:10 AM) Bronwyn Biro, Zhitian Zhang, Mo Chen and Angelica Lim Pose Forecasting in the SFU-Store-Nav 3D Virtual Human Platform [paper]
20:10 - 20:20 (11:10 - 11:20 AM) An Le, Philipp Kratzer, Simon Hagenmayer, Marc Toussaint and Jim Mainprice Hierarchical Prediction and Planning for Human-Robot Collaboration [paper]
20:20 - 20:30 (11:20 - 11:30 AM) Wenjie Yin, Hang Yin, Danica Kragic Jensfelt and Mårten Björkman Long-term Human Motion Generation and Reconstruction Using Graph-based Normalizing Flow [paper]
20:30 - 20:40 (11:30 - 11:40 AM) Lingfeng Sun, Masayoshi Tomizuka and Wei Zhan Multi-Style Human Motion Prediction and Generation via Meta-Learning [paper]