search

facebookresearch / OccupancyAnticipation

This repository contains code for our publication "Occupancy Anticipation for Efficient Exploration and Navigation" in ECCV 2020.
MIT License
76 stars 26 forks source link

Question in updating collision maps in ans.py. #21

Closed JeremyLinky closed 3 years ago

JeremyLinky commented 3 years ago

The t2 is equal to "global_pose[i, 2].item() - math.pi / 2", and then I visualize the collision map, finding that the collision you set is in front of the "global_pose[i, 2].item() - math.pi / 2" instead of the direction of agent's pose, which is "global_pose[i, 2].item()". Can you explain that? Much appreciate!!

srama2512 commented 3 years ago

There are two coordinate systems involved here.

(1) For the purposes of tracking the agent's 2D pose in the world map, we assume the following conventions: agent starts at the map center --- (0, 0) --- with X as forward, Y as rightward, and heading as positive starting from X to Y.

(2) We also use the standard image indexing for addressing the 2D map and for setting subgoals for planning. Here, the origin is the top-left corner, i.e., the agent starts at (H/2, W/2) where the map size is (H, W), and X is rightward, Y is downward, and heading is positive from X to Y.

In the line you are looking at, we basically convert the agent heading from coordinate system (1) to (2). 

                x1, y1 = asnumpy(self.states["prev_map_position"][i]).tolist()
                x2, y2 = asnumpy(self.states["curr_map_position"][i]).tolist()
                t2 = global_pose[i, 2].item() - math.pi / 2
JeremyLinky commented 3 years ago

There are two coordinate systems involved here.

(1) For the purposes of tracking the agent's 2D pose in the world map, we assume the following conventions: agent starts at the map center --- (0, 0) --- with X as forward, Y as rightward, and heading as positive starting from X to Y.

(2) We also use the standard image indexing for addressing the 2D map and for setting subgoals for planning. Here, the origin is the top-left corner, i.e., the agent starts at (H/2, W/2) where the map size is (H, W), and X is rightward, Y is downward, and heading is positive from X to Y.

In the line you are looking at, we basically convert the agent heading from coordinate system (1) to (2).

                x1, y1 = asnumpy(self.states["prev_map_position"][i]).tolist()
                x2, y2 = asnumpy(self.states["curr_map_position"][i]).tolist()
                t2 = global_pose[i, 2].item() - math.pi / 2

Thanks for your reply!