• On-site notes
• Speakers
• Program
• Organizers
• Call for papers
• Accepted Papers

Robot learning traditionally relies on applications of machine learning methods via the induction paradigm, i.e. where the agent extrapolates a policy in the form of a condensed, unified model given a set of training data. Under this formulation, training data is therefore discarded at test time, and the agent may no longer explicitly reason on past experience to inform its current action. In other words, the agent is not allowed to remember previous situations.
However, we observe that humans are able to conveniently recall past experiences to solve decision making problems, and notice a close connection with recent trends towards in-context learning, meta-learning, and retrieval-augmented inference in the Reinforcement Learning (RL) and Imitation Learning (IL) literature.

In this workshop, we investigate how robot learning algorithms can be designed for robotic agents to explicitly retrieve, reason, attend, or leverage on past experience to quickly adapt to previously unseen situations or bootstrap their learning.
Recent advances in the field allow agents to remember by framing the problem as in-context learning and meta-learning, e.g. by incorporating memory in the form of explicit, contextual input to a policy. Similarly, recent works propose retrieval mechanisms (retrieval-augmented policies) for agents to conveniently reason about the current action based on previous experience.

We then raise the question: "Should robots know, or remember how to act?"

Our objective is to understand the interconnection between the fields of in-context learning, meta-learning, transductive and retrieval-augmented inference for robot learning applications. In this context, we aim to discuss about recent contributions, challenges, opportunities and novel directions for future research.