Jaxnav docs & improvements #112
Conversation
| The default map is square robots of width 0.5m moving within a world with grid based obstacled, with cells of size 1m x 1m. Map cell size can be varied to produce obstacles of higher fidelty or robot strucutre can be changed into any polygon or a circle. | ||
|
|
||
| We also include a map which uses polygon obstacles, but note we have not used this code is a while so there may well be issues with it. |
|
|
||
| ### Observation space | ||
| By default, each robot recieves 200 range readings from a 360-degree arc centered on their forward axis. These range readings have a max range of 6m but no minimum range and are discritised with a resultion of 0.05 m. Alongside these range readings, each robot recieves their current linear and angular velocities along with the direction to their goal. Their goal direction is given by a vector in polar form where the distance is either the max lidar range if the goal is beyond their "line of sight" or the actual distance if the goal is within their lidar range. There is no communication between agents. |
There was a problem hiding this comment.
"discritised" -> "discretised"
There was a problem hiding this comment.
Also "recieves" -> "receives"
There was a problem hiding this comment.
the dyslexia do hit hard
| map_collisions = self._check_map_collisions(new_pos, new_theta, agent_states.map_data)*(1-agent_states.done).astype(bool) | ||
| agent_collisions = self._check_agent_collisions(jnp.arange(agent_states.pos.shape[0]), new_pos, agent_states.done)*(1- agent_states.done).astype(bool) | ||
| map_collisions = jax.vmap(self._map_obj.check_agent_map_collision, in_axes=(0, 0, None))(new_pos, new_theta, agent_states.map_data)*(1-agent_states.done).astype(bool) | ||
| # map_collisions = self._check_map_collisions(new_pos, new_theta, agent_states.map_data)*(1-agent_states.done).astype(bool) |
| @@ -708,11 +708,6 @@ def make_jaxnav_singleton_collection(collection_id: str, **env_kwargs) -> Tuple[ | |||
| "NarrowChicane2b", | |||
| "Chicane4", | |||
| ], | |||
| "new": [ | |||
There was a problem hiding this comment.
Why did you remove these tests? Ideally tests should be fixed unless you are removing a feature
There was a problem hiding this comment.
oh these were just used during development so tidying them into a clearer set
| @@ -330,6 +342,75 @@ def _passable(grid, posa, posb): | |||
|
|
|||
| def scale_coords(self, x): | |||
| return x / self.cell_size | |||
|
|
|||
| def dikstra_path(self, map_data, pos1, pos2): | |||
There was a problem hiding this comment.
Did you implement dijkstra's algorithm in Jax 😆
absolutely nuts good work
There was a problem hiding this comment.
Would be good to describe how this works in a high level comment
There was a problem hiding this comment.
at 11pm on a sunday yes haahah, Reviewer 4 asked for environment metrics..!
| @@ -569,6 +651,44 @@ def _checkPointRect(x, y, rx, ry): | |||
| inside = _checkPolyWithinRect(sides, grid_idx[1], grid_idx[0]) | |||
| return inside & map_grid[grid_idx[0], grid_idx[1]] & valid_idx | |||
|
|
|||
| @partial(jax.jit, static_argnums=[0]) | |||
| def check_all_agent_agent_collisions(self, agent_positions: chex.Array, agent_theta: chex.Array, agent_coords=None) -> chex.Array: | |||
There was a problem hiding this comment.
Also would be good to provide more detail for people (like) me who have no idea what the separating axis theorem is.
There was a problem hiding this comment.
yea very fair point :)
No description provided.