External Benchmarks/Tasks#

These are tasks and external libraries contributed by the community, built in ManiSkill. The document here has both a high-level overview/list of each project in a table as well as detailed task cards with video demonstrations after.

Task Table#

Table of all tasks/environments in this category. Task column is the environment ID, Preview is a thumbnail pair of the first and last frames of an example success demonstration. Max steps is the task’s default max episode steps, generally tuned for RL workflows.

Task

Preview

Dense Reward

Success/Fail Conditions

Demos

Max Episode Steps

ManiSkill-HAB
ManiSkill-HAB

✅

✅

✅

Full Task: 1200-2000

Subtasks: 200

ManiSkill-HAB#

Paper | Website | Code | Models | Dataset

dense-reward sparse-reward demos

Task Card

Task Description:

ManiSkill-HAB contains 3 long-horizon tasks:

  • TidyHouse: Move 5 target objects to different open receptacles (e.g. table, counter, etc).

  • PrepareGroceries: Move 2 objects from the opened fridge to goal positions on the counter, then 1 object from the counter to the fridge.

  • SetTable: Move 1 bowl from the closed drawer to the dining table and 1 apple from the closed fridge to the same dining table.

These tasks are split into 4 subtasks: Pick Place, Open, and Close.

Randomizations:

  • Target object/articulation/goal (depending on subtask)

  • Base x/y position and z rotation, arm and torso qpos noise

Success Conditions:

  • Pick: Object grasped, end-effector at rest position, robot static

  • Place: Object within 15cm of goal position, end-effector at rest position, robot static

  • Open: Articulation open, end-effector at rest position, robot static

  • Close: Articulation closed, end-effector at rest position, robot static

Fail Conditions:

  • Robot exceeds cumulative collision force threshold (5000N for Pick, 7500N for Place, 10,000N for Open/Close)

Citation

@article{shukla2024maniskillhab,
	author		 = {Arth Shukla and Stone Tao and Hao Su},
	title        = {ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks},
	journal      = {CoRR},
	volume       = {abs/2412.13211},
	year         = {2024},
	url          = {https://doi.org/10.48550/arXiv.2412.13211},
	doi          = {10.48550/ARXIV.2412.13211},
	eprinttype   = {arXiv},
	eprint       = {2412.13211},
	timestamp    = {Mon, 09 Dec 2024 01:29:24 +0100},
	biburl       = {https://dblp.org/rec/journals/corr/abs-2412-13211.bib},
	bibsource    = {dblp computer science bibliography, https://dblp.org}
}