Task Suite¶

The current paper and repository define 12 representative aerial manipulation tasks. They are grouped by the interaction structure that the policy and controller must handle.

Task summary¶

Category	Task	Terminal success	Intermediate progress
Instantaneous interaction	Press Button	button joint passes the pressed threshold	none
Instantaneous interaction	Peg-in-Hole	peg penetrates the opening within lateral bounds	none
Object transport	Frame Assembly	frame is aligned and placed on all pegs	lifted; placed
Object transport	Cabinet Pick-and-Place	can is placed stably on the drawer	door opened; can lifted; placed
Object transport	Lemon Harvesting	lemon is released inside the container	grasped/detached; delivered/released
Object transport	Toss Ball	released ball lands in the container under the base-position constraint	released; delivered
Articulated or constrained contact	Rotate Valve	shaft reaches the target rotation	engaged; rotated
Articulated or constrained contact	Push Slider	slider reaches its rail target	engaged; pushed
Articulated or constrained contact	Pull Lever	lever exceeds its target angular displacement	engaged; pulled
Articulated or constrained contact	Open Door	hinge exceeds the open threshold	engaged; opened
Articulated or constrained contact	Wipe Window	all stains are removed under contact constraints	per-stain removal
Articulated or constrained contact	NDT	stable contact is maintained at the inspection region	none

The registered environment family for each task contains multiple robot, action, and controller variants. Run python scripts/environments/list_envs.py for the authoritative current list.

Instantaneous interaction¶

Press Button¶

The end effector must approach a randomized button and drive its prismatic joint past the success threshold. Randomization includes wall pose and texture plus button pose, size, and color.

Peg-in-Hole¶

The peg tip must pass the wall surface while remaining inside the hole's lateral bounds. Wall and hole geometry, pose, texture, size, and color can vary.

Object transport¶

Frame Assembly¶

The robot lifts a frame, aligns it with a multi-peg fixture, and places it onto the pegs. Frame and peg initial poses are randomized.

Cabinet Pick-and-Place¶

The robot opens a cabinet door, lifts a can, and places the can stably on the drawer. Object pose, door friction, and payload mass can vary.

Lemon Harvesting¶

The robot grasps and detaches a lemon, transports it, and releases it into a container. Wall appearance and fruit locations are randomized.

Toss Ball¶

The robot must release a ball with sufficient momentum to reach a container while respecting a base-position constraint. This is the suite's explicitly dynamic transport task.

Articulated objects and constrained contact¶

Rotate Valve¶

The end effector engages a valve and rotates its shaft to the target. Wall and valve pose, texture, friction, and color can vary.

Push Slider¶

The end effector engages a slider and pushes its prismatic joint along a rail. Slider pose, friction, color, and surrounding wall conditions can vary.

Pull Lever¶

The robot engages a lever and pulls it through a target angular range while rejecting the reaction wrench.

Open Door¶

The end effector engages a door and drives the hinge past the open threshold. Door pose, texture, and hinge damping can be randomized.

Wipe Window¶

The robot removes stains by maintaining sufficient force and proximity while following the surface. Window pose, stain locations, and friction can vary.

NDT¶

The non-destructive-testing task requires stable contact at a randomized inspection point on a building surface for a fixed duration.

Randomization and comparability¶

Task configs expose task-specific randomization rather than one global difficulty switch. When comparing policies or controllers, record the task ID, seed, environment count, episode horizon, disturbance flags, and any config overrides. A result is comparable only when those conditions match.