Installation

Follow this guide to install UV.

The AI2-THOR simulation renders with Vulkan and needs the Vulkan loader (libvulkan.so.1): sudo apt-get install libvulkan1, or conda install -c conda-forge libvulkan-loader and add the environment's lib folder to LD_LIBRARY_PATH.

Running it

To run the entrypoint, run the following command.

 uv run -m semistaticsim.groundtruth.main

--config-name adaptive_exp runs the adaptation experiment instead, where the objects' routine changes at hour 504.

You can set the target amount of simulation time to collect as well as the scan size. The bigger the scan size, the higher the RAM usage and jitting time, but it has the potential to go faster if you want to collect a lot of simulation time.

After collecting the data, plot the groundtruth of the first pickupable across all its valid receptacles:

uv run -m semistaticsim.groundtruth.viz

To run the ai2thor simulation, you can use

 uv run -m semistaticsim.keyboardcontrol.main_skillsim

Data paths

Both stages write under generated_data/ in the working directory. Set data_dir to use another folder, with the same value for both stages:

uv run -m semistaticsim.groundtruth.main procthor_index=153 data_dir=/path/to/data
uv run -m semistaticsim.keyboardcontrol.main_skillsim mode=auto index=153 data_dir=/path/to/data

How-to

Features are based around varying the level of scene-to-scene semantic transfer. Every simulator step, some dt time elapses. When the duration_left reaches 0, the transition model selects the next receptacle that the object will transition to. Then, the duration model sampels the amount of tiem that the objet will spend in that new receptacle.

Duration model:

1. even is "evenly spread duration of all steps"
2. instant is "spend NO time at this place, immediately transition at the next step" (this is what flowmaps currently has in your 2D simulator)
3. deterministic is "randomly split the day among all steps"
4. gaussian is the same as deterministic with some gaussian noise

Transition model:

1. "fixed_canonical": object cycles down a list of fixed receptacles
2. "fixed_0.1_0.9": object has 10% chance of staying put, 90% chance of going to the next receptacle in cycle (this is what the 2D flowmaps simulator has)
3. "uniform_no_diag": fully uniform transition matrix
4. "uniform_full": fully uniform transition matrix
5. "location_weighted_uniform_no_diag": uniform transition matrix weighted by the ProcThor receptacle prior
6. "location_weighted_uniform_full": uniform transition matrix weighted by the ProcThor receptacle prior

Preliminary scan experiments

1. SCANSIZE 10: ~33.5 it/s : 335 steps/s
2. 100 : 33 : 3300

1000 : eta 50min

1000 : cpu: 3s/it; cuda is same!

To build a package for PyPi

# bump the version in the .toml

uv build

# delete the old build in ./dist

uv publish