Task Sampling
Generate fresh instances of existing tasks with randomized elements for variety and robustness testing.
Getting Started
Clone 2026-challenge-task-instances into gm.DATA_PATH:
Sampling Workflow
Step 1: Review BDDL and Generate the JSON Template
Pick a task,review the bddl definition under bddl3/bddl/activity_definitions/TASK_NAME/problem_0.bddl. Make sure the defintion is reasonable. In particular watch out for wildcard expansions.
Then, generate a JSON template for your task:
The script will interactively prompt you to:
- Scene: choose from
house_double_floor_lower,house_double_floor_upper,house_single_floor, or enter a custom scene name. - Models: for each required synset and category, choose one or more model IDs from those available on disk. A link to the synset page on the BEHAVIOR Knowledgebase (e.g. https://behavior.stanford.edu/knowledgebase/synsets/ashcan.n.01.html) is printed alongside each prompt to help you browse available models.
The script writes the completed entry directly to datasets/2026-challenge-task-instances/metadata/task_custom_lists.json. The result looks like:
"picking_up_trash": {
"room_types": [
"living_room",
"kitchen"
],
"house_double_floor_lower": {
"whitelist": {
"can__of__soda.n.01": {
"can_of_soda": {
"itolcg": null,
"lugwcz": null,
"opivig": null
}
},
"ashcan.n.01": {
"trash_can": {
"wkxtxh": null
}
}
},
"blacklist": {}
}
}
Step 2: Sample Task-Related Objects (TRO)
It is highly recommended to run this command with -m pdb, so it will stop at the error during sampling and you can debug interactively to see what's wrong.
The scene is read automatically from task_custom_lists.json. After this command, you should see 2 files generated under datasets/2026-challenge-task-instances/scenes/SCENE_NAME/json: house_double_floor_lower_task_picking_up_trash_0_0_template-partial_rooms.json (intermediate) and house_double_floor_lower_task_picking_up_trash_0_0_template.json (postprocessed, with full scene objects merged in).
Step 3: Generate Instances
Randomly generate 1 instances for your task:
python OmniGibson/scripts/sampling/multiply_b1k_tasks.py --partial_save --start_idx 1 --end_idx 1 -t TASK_NAME
After this step, a folder named house_double_floor_lower_task_picking_up_trash_instances appears under datasets/2026-challenge-task-instances/scenes/SCENE_NAME/json, containing files named like house_double_floor_lower_task_picking_up_trash_0_1_template-tro_state.json.
Step 4: Presample Robot Poses
Step 5: Register New Task
Step 6: Update Task Misc
Assign an ID to the task (also update the google sheet!), then put a new entry in 2026-challenge-task-instances/metadata/B100_task_misc.csv. Take a look at the floor plan, and put in the task relevant rooms in the entry. Think of it as "what are the minimal set of rooms that is required for the robot to complete the task as if it's in a fully-loaded scene?"
Note that the rooms should include not only the rooms that contains the robot and TROs, but also those that connects them in between (e.g. corridors), as well as any room that could be in sight of the robot. Take house single floor as an example, for a task that only requires kitchen_0, we will need to load in the following 7 rooms: - corridor_0 - dining_room_0 - entryway_0 - garden_0 - kitchen_0 - living_room_0 - living_room_1
Step 7: Verify Task Viability
Prepare the joylo device, and run the following commands:
You should be able to complete the task without major bottlenecks. Watch out for any issues during teleoperation. Here are some examples:
- Cannot complete the task (e.g. not able to navigate to a room because of narrow corridor)
- Major artifacts / bad appearances in the scenes or objects.
- The task requires a lot of effort to complete (e.g. need to pick something up from a very high cabinet).
- The tasks induces unavoidable collisions between robot and the environment to complete (e.g. robot can't pick up food from the oven without colliding with the door)
- Other unreasonable behavior during teleoperation (e.g. object is too heavy to pickup, door is very hard to open, etc.)
If any of the above happens, either redo the previous sampling steps while fixing bugs, or if it's unfixable, discard the task and restart with another task.
After teleoperation succeeds, you should see a hdf5 file at HDF_PATH. Run the following replay script, which will generate the video and QA result JSON file:
Check the QA json file output. All QA should pass unless for task-specific reasons (e.g. one failed grasp is allowed for turning_on_radio because the gripper needs to close to poke the button). Also check mp4 to make sure the replay visual looks reasonable. For example, you shouldn't see a whole in the ground, which might indicates you forgot to put one room in the task misc csv.
If all outputs seems reasonable, share the generated MP4 file and QA result with the team for review.
Step 8: Generate the rest of the 300 instances
Run the multiply script again, this time with index 2 to 300, and then sample robot poses, then update task yaml:
python OmniGibson/scripts/sampling/multiply_b1k_tasks.py --partial_save --start_idx 2 --end_idx 300 -t TASK_NAME -s SCENE_NAME
Step 9: Prepare all files and submit PR
After the task design is finalized, create a seperate branch in 2026-challenge-task-instances, commit the files created:
- two seed instance json files: `0_0_template.json`, `0_0_template-partial_rooms.json`
- 300 task intance files under
- updated `task_custom_list.json` and `available_tasks.yaml`
Watch out for merge conflicts from main, which will most likely happen on task_custom_list.json and available_tasks.yaml.
Submit a PR and tag the team for review.