petkopetkov/longhorizon-subtask-decomposition

Long-horizon subtask decomposition

Given the initial frame of a tabletop manipulation scene and a compound instruction, produce the ordered list of single pick-and-place subtasks that completes it. One row per scene.

The frames and labels are taken from three long-horizon composition suites built on SimplerEnv (WidowX/Bridge), RobotArena (WidowX) and RoboLab (Franka). Each row's frame is the reset observation the orchestrator is actually shown at inference: the same camera, the same deterministic layout, and the same instruction string an evaluation run would pass. Nothing is re-derived or re-rendered for presentation.

Fields

field	meaning
id	<benchmark>/<scene>
benchmark	simplerenv_comp, robotarena_comp, robolab_comp
instruction	the compound instruction the policy reads
subtasks	ground-truth ordered subtask list
n_goals	goals the scene scores. Equal to len(subtasks) except for RoboLab relay tasks, where each goal needs two placements (via an intermediate receptacle)
ordered	whether the goals must be completed in the listed order
distractors	objects present but not part of any goal
family	RoboLab task family (single_dest, split, relay, transfer, …)
image_main	the reset frame, as the orchestrator sees it
image_wrist	wrist camera; present for RoboLab only, null elsewhere

Scoring

Three properties make naive scoring wrong, so an evaluation should:

Match subtasks as a multiset, not a set. 30 of the SimplerEnv scenes require the same subtask more than once — all10_plate holds four separate carrots and needs put the carrot on the plate four times. Set semantics score a plan that names each object once as complete.

Treat the two phrasings of an elevated-source goal as equivalent. A goal whose object starts on a plate can be written move the carrot from the plate to the blue towel or put the carrot on the blue towel. The suites' plan caches and their environments disagree about which to use for the same goal, and both are correct, so normalise the three-slot form onto the two-slot one before comparing.

Ignore order unless ordered is set. Most scenes admit any permutation.

Granularity needs no separate metric: splitting one pick-and-place into a grasp and a place shows up as a positive length error and as unparsable lines, both of which fall out of the above.

Caveats

• Plan lengths run 1–10 goals; the long tail is where decomposition errors concentrate.
• One frame per scene, deliberately. The orchestrator decomposes once per scene from a single observation and reuses that plan for every episode, so per-episode layout variation is not part of the task being measured.