News
Recording is available at this link!
The in-person workshop is at 414 CEPSR, Columbia University Campus
The in-person poster session is at the CS Lounge, Mudd Building 1st Floor
Gather.Town for poster session (11:00am - 12:30pm EST): https://app.gather.town/app/CPO3rXXGahiPKGZi/RSS2022-ILWS
Abstract
Imitation learning is a promising paradigm to teach autonomous agents to perform sequential decision making tasks by allowing them to learn from a set of task demonstrations provided by an expert. There has been enormous progress in developing algorithmic tools for imitation learning, and advancing our theoretical understanding of the guarantees provided by such algorithms. However, there are other aspects that are important to scale the imitation learning paradigm to more complex, real-world scenarios. Commonly overlooked aspects include the manner in which data is collected, the quality of the demonstrator, evaluating the quality of data and the performance of algorithms across a broad range of tasks, and how the design of data collection systems and algorithms can influence each other.
Topics of Interest
This workshop seeks to bring together researchers and practitioners to discuss these overlooked aspects of imitation learning. We categorize the aspects into the following four high-level topics:
Systems: Modern imitation learning algorithms can learn from demonstration data that is diverse in format (e.g., raw sensory streams) and source (e.g., suboptimal expert, policy-in-the-loop). This topic focuses on new systems and mechanisms for collecting demonstration data, including (but not limited to) interfaces for operating / tele-operating robots, mechanisms for on-demand expert intervention, and systems for collecting human-robot collaboration data.
Data: Who should be the demonstrators and for what algorithms? This topic focuses on the interplay of data sources and learning algorithms, e.g., the effects of learning from imperfect demonstrations, from multiple experts, and from mixed human-generated data and machine-generated data.
Tasks and Domains: Imitation learning systems and algorithms are often coupled with tasks and domains. Insights from studying one domain, e.g., manipulation, may not transfer to another, e.g., autonomous driving. On the other hand, domain knowledge can often be exploited to drastically improve imitation learning systems and algorithms. For this topic, we encourage contributions that highlight imitation learning efforts that either (a) transcend different tasks and domains or (b) effectively leverage domain knowledge to improve a general-purpose system or algorithm.
Submission Instructions and Dates
We invite submissions that are 4-8 pages long (no limit on references or supplementary material) that have not been accepted at an archival venue. Submitted papers should follow the RSS 2022 paper format (see this link) and should be anonymized for review.
Submission portal link (CMT): https://cmt3.research.microsoft.com/RSSILWS2022/
Important Dates:
May
713, 2022 Submission Deadline (11:59pm PT)May
2325, 2022 Author NotificationJune 12, 2022 Camera Ready Submission
June 27, 2022 Workshop
Hybrid Format
Live streaming: We plan to stream the workshop proceedings over a Zoom meeting room in order to try and make it as easy as possible for participants (both local and remote) to join and participate. All speakers and spotlight presenters are expected to participate in real time, with an option of playing pre-recorded videos for virtual presenters who cannot present at their time slot due to time difference.
Q & A for virtual attendees: We will also have an online question inbox that we will be monitoring throughout the day, to allow participants to ask questions or general discussion points to any of our speakers. These will be answered during dedicated Q&A sections of each speaker’s talk.
Talk format: To further encourage participation and interaction, we have encouraged our speakers to limit their talks to 20 minutes, allowing for at least 10 minutes of discussion time for each speaker. We believe that the most important function of this workshop is to enable participants to discuss different perspectives on commonly overlooked topics in imitation learning. Along these lines, we also encourage remote participants to participate in our poster sessions. We have provided ample time for the poster sessions to encourage discussion.
Hybrid poster session: We plan to facilitate a hybrid poster session over GatherTown. Each poster stand should be equipped with a laptop or a tablet that’s capable of accessing GatherTown. Virtual presenters will present through the GatherTown interface to virtual attendees. In-person attendees will also be able to interact with virtual presenters through the laptop / tablet at the poster stand. Conversely, physical presenters should be able to interact with virtual attendees through the laptop.