Closed dmcconachie-tri closed 7 months ago
scene_model: Benevolence_1_int
activity_name: clean_a_kitchen_table
Note there are extra errors in the middle of the log:
scene_model: Beechwood_0_int
activity_name: make_a_salad
scene_model: Rs_int
activity_name: prepare_sea_salt_soak
Oddities:
Hi @dmcconachie-tri ,
Huge thanks for taking the time to compile these bug reports. Our team is taking a look into it; we should hopefully have fixes ready by tomorrow!
Hi @dmcconachie-tri ,
Apologies for the delay. We have a new release out (0.2.1) that fixes some of your issues. Concretely:
scene_model: Benevolence_1_int
activity_name: clean_a_kitchen_table
I was unable to deterministically reproduce this error on my machine. It occurred the first time I ran but did not show up during subsequent runs. This is a bit concerning, as it seems to suggest potentially an NVIDIA / Omniverse caching issue of some sort. We are hoping it can be reproduced potentially on another setup (@cgokmen )
scene_model: Beechwood_0_int
activity_name: make_a_salad
I was able to reproduce the segfaults deterministically. We had similar issues with IsaacSim v2022.2.1 (stochastically) but had never seen it occur with 2022.2.0. We're hoping that the impending NV release v2023.1.0 will resolve this issue.
scene_model: Rs_int
activity_name: prepare_sea_salt_soak
This (as well as other problematic tasks with similar object-through-floor) has been resampled. Moreover, salt should no longer explode out of the container.
Regarding salt particle sizes, it is a known limitation of the NVIDIA Omniverse particle systems. We could reduce the diameter of the particles, but physics stability is likely to decrease, especially at the default physics timestep size (1/60). In general, we've tried to limit our particles from being less than 1-2cm in diameter for this reason, at the expense of a bit less realism. Currently there's no explicit API that allows you to modify the particle size at runtime, but you can always go to the hyperparameter specification (in <DATASET>/systems/<system_name>/metadata.json
) and directly modify the size there.
As an aside, there's also been other problematic tasks (not mentioned explicitly in this thread) and issues that have been fixed as well. Apologies for the issues; again, huge thanks for tagging these bugs and being patient with the development. We're planning on having another mini-release soon once NVIDIA releases their new IsaacSim version -- hopefully within the next 3 weeks or so!
May I ask where I can run the Behavior 1K tasks?
It seems these problems haven't been resolved... I install Omnigibson with docker on linux without GUI as the link OmniGibson Documentation. However when I run python -m omnigibson.examples.environments.behavior_env_demo
(the only change is to enable USE_GPU_DYNAMICS), I also meet the error:
2023-11-26 11:58:24 [1,146,503ms] [Error] [omni.kit.manipulator.prim.model] Traceback (most recent call last): File "/isaac-sim/kit/extscore/omni.kit.manipulator.prim/omni/kit/manipulator/prim/model.py", line 1492, in _update_transform_from_prims_async pivot_prim_path = self._pivot_prim.GetPath() AttributeError: 'NoneType' object has no attribute 'GetPath'
I believe all of the above issues are fixed in our current development branch - closing this issue, please keep an eye out for our upcoming release.
I'm trying to experiment with the Behavior 1K tasks and I'm running into various errors depending on the task that I attempt ranging from
to
and straight up segfaults.
In all instances these errors have been replicated both in install from source and docker versions.
Replication steps:
sudo ./run_docker.sh <path_to_data>
(note no-h
, though presumably not relevant)/omnigibson-src/omnigibson/macros.py
to enableUSE_GPU_DYNAMICS
based on instructions/error message when it is not enabled./omnigibson-src/omnigibson/configs/fetch_behavior.yaml
to use a particular scenepython -m omnigibson.examples.environments.behavior_env_demo
git log -n1
from my local (non-docker) checkoutExample git diff from upstream/main
``` diff --git a/omnigibson/configs/fetch_behavior.yaml b/omnigibson/configs/fetch_behavior.yaml index 84da35e3..28afb218 100644 --- a/omnigibson/configs/fetch_behavior.yaml +++ b/omnigibson/configs/fetch_behavior.yaml @@ -7,7 +7,7 @@ render: scene: type: InteractiveTraversableScene - scene_model: Rs_int + scene_model: Benevolence_1_int trav_map_resolution: 0.1 trav_map_erosion: 2 trav_map_with_objects: true @@ -50,7 +50,7 @@ objects: [] task: type: BehaviorTask - activity_name: prepare_sea_salt_soak + activity_name: clean_a_kitchen_table activity_definition_id: 0 activity_instance_id: 0 predefined_problem: null diff --git a/omnigibson/macros.py b/omnigibson/macros.py index 6538479e..fe76b9a6 100644 --- a/omnigibson/macros.py +++ b/omnigibson/macros.py @@ -40,7 +40,7 @@ gm.SHOW_DISCLAIMERS = False # Whether to use omni's GPU dynamics # This is necessary for certain features; e.g. particles (fluids / cloth) -gm.USE_GPU_DYNAMICS = False +gm.USE_GPU_DYNAMICS = True # Whether to use high-fidelity rendering (this includes, e.g., isosurfaces) gm.ENABLE_HQ_RENDERING = False ```Full error messages in followup posts.