Episodes: The Structured Unit of Robot Learning
Raw trajectories and clips are the wrong granularity for learning from physical experience. We define the episode — goal, context, objects, actions, state changes, failures, recoveries, interventions, and outcome — and argue why it is a more compact and reusable learning signal than disconnected demonstrations.
Robot learning is not bottlenecked by model size alone. It is bottlenecked by the quality of physical experience — and by the form that experience takes when we try to learn from it.
The problem with fragments
Today, robotics data exists as raw videos, teleoperation traces, human demonstrations, simulation rollouts, failures, recoveries, and deployment logs. These are not native learning units. They are fragments.
Scaling fragments naively produces more data, not necessarily more intelligence. A longer video or another trajectory adds volume without adding structure — and structure is what a model needs in order to generalize from few examples.
Defining the episode
An episode is the structured unit of robot learning. It captures the goal, context, objects, actions, state changes, failures, recoveries, interventions, and outcome of a single interaction.
Where a trajectory is a sequence of low-level states and actions, an episode is a self-contained account of what was being attempted, what happened, what went wrong, how it was corrected, and whether it succeeded.
Why episodes are more efficient
Episodes are compact. They discard redundancy and keep the information that explains an outcome, so each one carries more signal per unit of data.
Episodes are reusable. Because they encode goal and context explicitly, they can be retrieved and recombined across tasks and embodiments rather than memorized in isolation.
Episodes make failure useful. Failures, recoveries, and interventions — usually discarded as noise — become first-class signal, and they are often the most informative part of an interaction.
What we are building
Episode Intelligence is building the data-efficiency layer for robot learning: the system that turns raw physical experience into episodes, and episodes into reusable learning signal.
Our goal is simple to state and hard to achieve — to make robots learn more from fewer real-world interactions.