godot_rl_JumperHard / Player.gd
edbeeching's picture
edbeeching HF staff
Upload . with huggingface_hub
b2f36c3
raw
history blame
6.65 kB
extends CharacterBody3D
const MOVE_SPEED = 12
const JUMP_FORCE = 30
const GRAVITY = 0.98
const MAX_FALL_SPEED = 30
const TURN_SENS = 2.0
const MAX_STEPS = 20000
@onready var cam = $Camera3D
var move_vec = Vector3()
var y_velo = 0
var needs_reset = false
# RL related variables
@onready var end_position = $"../EndPosition"
@onready var raycast_sensor = $"RayCastSensor3D"
@onready var first_jump_pad = $"../Pads/FirstPad"
@onready var second_jump_pad = $"../Pads/SecondPad"
@onready var robot = $Robot
var next = 1
var done = false
var just_reached_end = false
var just_reached_next = false
var just_fell_off = false
var best_goal_distance := 10000.0
var grounded := false
var _heuristic := "player"
var move_action := 0.0
var turn_action := 0.0
var jump_action := false
var n_steps = 0
var _goal_vec = null
var reward = 0.0
func _ready():
raycast_sensor.activate()
reset()
#func _process(_delta):
# if _goal_vec != null:
# DebugDraw.draw_line_3d(position, position + (_goal_vec*10), Color(1, 1, 0))
func _physics_process(_delta):
#reward = 0.0
n_steps +=1
if n_steps >= MAX_STEPS:
done = true
needs_reset = true
if needs_reset:
needs_reset = false
reset()
return
move_vec *= 0
move_vec = get_move_vec()
#move_vec = move_vec.normalized()
move_vec = move_vec.rotated(Vector3(0, 1, 0), rotation.y)
move_vec *= MOVE_SPEED
move_vec.y = y_velo
set_velocity(move_vec)
set_up_direction(Vector3(0, 1, 0))
move_and_slide()
# turning
var turn_vec = get_turn_vec()
rotation.y += deg_to_rad(turn_vec*TURN_SENS)
grounded = is_on_floor()
y_velo -= GRAVITY
var just_jumped = false
if grounded and get_jump_action():
robot.set_animation("jump")
just_jumped = true
y_velo = JUMP_FORCE
grounded = false
if grounded and y_velo <= 0:
y_velo = -0.1
if y_velo < -MAX_FALL_SPEED:
y_velo = -MAX_FALL_SPEED
if y_velo < 0 and !grounded :
robot.set_animation("falling")
var horizontal_speed = Vector2(move_vec.x, move_vec.z)
if horizontal_speed.length() < 0.1 and grounded:
robot.set_animation("idle")
elif horizontal_speed.length() < 1.0 and grounded:
robot.set_animation("walk")
elif horizontal_speed.length() >= 1.0 and grounded:
robot.set_animation("run")
update_reward()
if Input.is_action_just_pressed("r_key"):
reset()
func get_move_vec() -> Vector3:
if done:
move_vec = Vector3.ZERO
return move_vec
if _heuristic == "model":
return Vector3(
0,
0,
clamp(move_action, -1.0, 0.5)
)
var move_vec := Vector3(
0,
0,
clamp(Input.get_action_strength("move_backwards") - Input.get_action_strength("move_forwards"),-1.0, 0.5)
)
return move_vec
func get_turn_vec() -> float:
if _heuristic == "model":
return turn_action
var rotation_amount = Input.get_action_strength("turn_left") - Input.get_action_strength("turn_right")
return rotation_amount
func get_jump_action() -> bool:
if done:
jump_action = false
return jump_action
if _heuristic == "model":
return jump_action
return Input.is_action_just_pressed("jump")
func reset():
needs_reset = false
next = 1
n_steps = 0
first_jump_pad.position = Vector3.ZERO
second_jump_pad.position = Vector3(0,0,-12)
just_reached_end = false
just_fell_off = false
jump_action = false
set_position(Vector3(0,5,0))
rotation.y = deg_to_rad(randf_range(-180,180))
y_velo = 0.1
reset_best_goal_distance()
func set_action(action):
move_action = action["move"][0]
turn_action = action["turn"][0]
jump_action = action["jump"] == 1
func reset_if_done():
if done:
reset()
func get_obs():
var goal_distance = 0.0
var goal_vector = Vector3.ZERO
if next == 0:
goal_distance = position.distance_to(first_jump_pad.position)
goal_vector = (first_jump_pad.position - position).normalized()
if next == 1:
goal_distance = position.distance_to(second_jump_pad.position)
goal_vector = (second_jump_pad.position - position).normalized()
goal_vector = goal_vector.rotated(Vector3.UP, -rotation.y)
goal_distance = clamp(goal_distance, 0.0, 20.0)
var obs = []
obs.append_array([move_vec.x/MOVE_SPEED,
move_vec.y/MAX_FALL_SPEED,
move_vec.z/MOVE_SPEED])
obs.append_array([goal_distance/20.0,
goal_vector.x,
goal_vector.y,
goal_vector.z])
obs.append(grounded)
obs.append_array(raycast_sensor.get_observation())
return {
"obs": obs,
}
func get_obs_space():
# typs of obs space: box, discrete, repeated
return {
"obs": {
"size": [len(get_obs()["obs"])],
"space": "box"
}
}
func update_reward():
reward -= 0.01 # step penalty
reward += shaping_reward()
func get_reward():
var current_reward = reward
reward = 0 # reset the reward to zero checked every decision step
return current_reward
func shaping_reward():
var s_reward = 0.0
var goal_distance = 0
if next == 0:
goal_distance = position.distance_to(first_jump_pad.position)
if next == 1:
goal_distance = position.distance_to(second_jump_pad.position)
#print(goal_distance)
if goal_distance < best_goal_distance:
s_reward += best_goal_distance - goal_distance
best_goal_distance = goal_distance
s_reward /= 1.0
return s_reward
func reset_best_goal_distance():
if next == 0:
best_goal_distance = position.distance_to(first_jump_pad.position)
if next == 1:
best_goal_distance = position.distance_to(second_jump_pad.position)
func set_heuristic(heuristic):
self._heuristic = heuristic
func get_obs_size():
return len(get_obs())
func zero_reward():
reward = 0
func get_action_space():
return {
"move" : {
"size": 1,
"action_type": "continuous"
},
"turn" : {
"size": 1,
"action_type": "continuous"
},
"jump": {
"size": 2,
"action_type": "discrete"
}
}
func get_done():
return done
func set_done_false():
done = false
func calculate_translation(other_pad_translation : Vector3) -> Vector3:
var new_translation := Vector3.ZERO
var distance = randf_range(12,16)
var angle = randf_range(-180,180)
new_translation.z = other_pad_translation.z + sin(deg_to_rad(angle))*distance
new_translation.x = other_pad_translation.x + cos(deg_to_rad(angle))*distance
return new_translation
func _on_First_Pad_Trigger_body_entered(body):
if next != 0:
return
reward += 100.0
next = 1
reset_best_goal_distance()
second_jump_pad.position = calculate_translation(first_jump_pad.position)
func _on_Second_Trigger_body_entered(body):
if next != 1:
return
reward += 100.0
next = 0
reset_best_goal_distance()
first_jump_pad.position = calculate_translation(second_jump_pad.position)
func _on_ResetTriggerBox_body_entered(body):
done = true
reset()