Long-horizon task

“Long” describes structure as well as duration

Some papers define a horizon by environment steps or elapsed time. Others use the term for tasks composed of dependent phases such as opening a drawer, retrieving an item and placing it elsewhere. A 2026 scoping review found substantial variation and treats temporally extended decision-making across dependent actions, skills or contacts as the common structure.

A task can therefore be difficult over a long horizon even when each local movement is simple. Failure in an early phase may make every later action impossible.

Execution and memory create different failures

Small pose, timing or contact errors can accumulate across repeated actions. Other tasks are partially observable: the current image may not reveal which objects were already handled or which instruction phase is active.

LongBench separates fully observable long execution from context-dependent tasks that require historical information. This distinction prevents every long-horizon failure from being attributed to missing memory when the real cause is unreliable physical execution.

Episode design changes what can be learned

CALVIN evaluates sequences of language-conditioned manipulation tasks rather than isolated single skills. Training data for such problems should preserve subtask order, intermediate object state, corrections, failures, recovery and the language or goal active at each phase.

Trimming episodes into successful short clips can remove the dependencies that define the task. Evaluation should report complete-sequence success and per-phase failure, with reset rules and maximum horizon stated explicitly.