presentation is here

Visual Navigation in indoor environment with Top-down Semantic Map.

by taking advantage of value iteration network, the action policy network is based on the value map generated.

Each Semantic Map has a size of H X W X C where H and W stands for height and width, and C is the number of object categories. Each cell on Semantic Map has a vector of object occupancies.

• Add functions in preprocess/patch_core.py to the House3D/House3D/core.py file under the Environment class.

• Add functions in preprocess/patch_house.py to the House3D/House3D/house.py file under the House class.

• Integrate the colormap.csv to the House class by adding line 7 to 13 to House Class. This also generates self.smap and self.smap_img to the Class.

• preprocess/genhouseinfo.py generates the necessary house information for local semantic map as well. Modify the HOUSEDIR, CONFIGFILEPATH and house_ids. run by python preprocess/genhouseinfo.py.

• preprocess/gensmap.py provides with a gensmap class and generates the local semantic map of 94 classes at every location given in a map.txt file previously generated (not included in the repo). change the house_ids and lmapszs to desired value and run by python preprocess/gensmap.py.

• get_tar_star_minsteps_aseq.py generates action sequency for every location in given map.txt file for each given target.

• res_vis.py gives a solution to generate all frames of learnt reward map and value map and local semantic map along the trajctory.

• src/navi_env.py interacts with the aforementioned generated data.

• src/multienv.py builds upon nav_env.py and interacts with multiple environments and enable multiple agents approaching different target (one agent one target).

• tf_code/nav_agent_release.py initialize training or testing.