Skip to content

Latest commit

 

History

History
54 lines (38 loc) · 4.62 KB

custom_agent.md

File metadata and controls

54 lines (38 loc) · 4.62 KB
Raw

Custom Agent

Please follow the instructions in this page if you want to submit your own agent.

Agent Implementation

The agent need to be class object with two callable functions: reset and choose_target

We provide an empty agent framework at policy/custom.py. To design your own agent, please reimplement function self.reset and self.choose_target

Reset Function

self.reset(): this function will be called before each episode. When it is called, agent will receive the following information. The information of all target objects is given in objects_info

Input Parameter Type Description Example
goal_objects List[str] List of target categories ["pen", "bowl", "lighter", "key"]
objects_info Dict Keys are category names. Values includes the value, waterproofness or inflammability {"pen":{"value": 1, "waterproof": 1}}

Choose Target Function

self.choose_target: this function will be called when the trial begins or agent receives the execution result of the last action. This function must return with one action, which will be executed by the environment. It has two dict-type inputs parameters, state and processed_input, described as the following:

state:

Key Value type Description
raw Dict Has 4 keys: rgb, depth, log_temp, and seg_mask. Their values are RGB signals, depth signals, environmental signals (log temperature for fire and water level for flood), and segmentation mask signals (unavailable if you use the perceptional version).
sem_map Dict Semantic map built by agent 2D observations + depth. Has 4 keys: height, explored, id, and others. Their values are height map, explored map (0 or 1), object id map, and other information.
goal_map numpy.array A map of target objects and agent itself. If the agent has spotted the targets, targets will be value 1 in this map. The agent itself will be value -2 in this map. The other elements are 0.
RL Dict Information used to train RL agent. Can be ignored here.

procesed_input:

This function should return a action for agent to execute, in the format of (action_name, params). Agent can either choose a high-level action or a low-level one. The available actions are:

  • Available high-level actions
Action name Params Description
walk_to Int (target id) Agent will try to walk to a target object
pick_up None or Int (target id) Agent will try to pick up a object (if not specified, agent will try to pick up the object agent just walked to)
drop None If agent has picked up an object, the agent will drop it (in wind task), or put it in its bag (fire or flood task)
explore None Agent will turn by 360 degrees, The observations during turning will update the semantic map of the agent
  • Available low-level actions

Currently only support 4 low-level actions. The params should follow the format of TDW replicant. Please refer the documentation of TDW Replicant for more details.

Action name Params Description
move_by Dict Agent will try to move by a certain distance
turn_by Dict Agent will turn by a certain degree
turn_to Dict Agent will turn to a given degree
reach_for Dict Agent will try to reach for a given position

To avoid cheating, the following keys in params are forbidden: max_distance, duration, scale_duration, and arrived_at.