Speakers
CoRL 2026 · Half-Day Workshop
Toward persistent autonomy: what robots and vision–language–action models should remember — and how memory should be learned, built, evaluated, and deployed.
About the Workshop
A robot in the physical world cannot treat every timestep as a fresh start — it may need to recall that a mug went into a cupboard, that a drawer was already checked, or that a grasp failed because the object slipped. By memory we mean the ability to store past experience and call it back when it is relevant, not just to keep recent observations in view — which matters most in partially observable, long-horizon tasks. Yet most robot foundation models — generalist policies, vision–language–action (VLA) models, and video world models — treat memory only shallowly: a window of recent context, a replay buffer, or retrieval over past observations.
This workshop works toward a shared vocabulary for robot memory, the failure modes it must address, and evaluation criteria that isolate what memory actually contributes.
What should memory mean for autonomous robots — and how should we learn, build, evaluate, and deploy it?
Speakers
Program
Open Problems
Beyond perception and policy execution, what should memory enable — tracking hidden state, avoiding repeated mistakes, recovering from failures, adapting to environments and users, and accumulating reusable knowledge?
Observations, language summaries, object states, maps, action traces, failures, preferences, or latent representations? And how can a robot retrieve the right memory under partial observability without surfacing stale or irrelevant entries?
How should robots compress repeated experience into reusable knowledge, preserve rare but important events, and revise or forget memories when environments, tasks, or users change?
Should memory be external, latent, structured, or end-to-end learned — and how should it interact with reactive control, long-horizon planning, prediction, and language-conditioned reasoning?
How can memories of failed grasps, unsuccessful searches, misunderstood instructions, or changing environments lead to improved future behavior rather than passive recall?
What benchmarks and metrics can distinguish genuine memory use from larger context windows, environment memorization, or task-specific shortcuts?
Call for Papers
Submission portal opens soon.
CoRL 2026 · Austin, TX, USA.
