team222 / env_3d.py
ylop's picture
Deploy 2M step LNN training with optimized GPU utilization
28dbd6d verified
raw
history blame
10.4 kB
import numpy as np
import gymnasium as gym
from gymnasium import spaces
import pybullet as p
from PyFlyt.core import Aviary
from PyFlyt.core.drones import QuadX
class Drone3DEnv(gym.Env):
metadata = {"render_modes": ["human"]}
def __init__(self, render_mode=None):
super().__init__()
self.render_mode = render_mode
# Constants
self.BOUNDS = np.array([[-5, -5, 0], [5, 5, 10]]) # x_min, y_min, z_min, x_max, y_max, z_max
self.TARGET_BOUNDS = np.array([[-4, -4, 1], [4, 4, 9]])
self.MAX_STEPS = 1000
self.WIND_SCALE = 1.0
# Initialize PyFlyt Aviary
self.start_pos = np.array([[0.0, 0.0, 1.0]])
self.start_orn = np.array([[0.0, 0.0, 0.0]])
self.aviary = Aviary(
start_pos=self.start_pos,
start_orn=self.start_orn,
drone_type="quadx",
drone_options=[{
"model_dir": "/home/ylop/Documents/drone go brr/Drone-go-brrrrr/Drone-go-brrrrr/my_models",
"drone_model": "cf2x_big"
}],
render=self.render_mode is not None
)
# Action Space: 4 motors (0..1)
self.action_space = spaces.Box(low=-1.0, high=1.0, shape=(4,), dtype=np.float32)
# Observation Space: 12 dim
self.observation_space = spaces.Box(low=-np.inf, high=np.inf, shape=(12,), dtype=np.float32)
self.wind_vector = np.zeros(3)
self.target_pos = np.zeros(3)
self.step_count = 0
# Visual IDs
self.wind_arrow_ids = []
self.target_visual_id = None
self.bound_box_ids = []
self.path_points = []
self.text_ids = []
def reset(self, seed=None, options=None):
super().reset(seed=seed)
self.aviary.reset()
self.step_count = 0
self.path_points = []
# Clear text
for uid in self.text_ids:
p.removeUserDebugItem(uid)
self.text_ids = []
# Disable PyBullet GUI sidebars but enable mouse
if self.render_mode == "human":
p.configureDebugVisualizer(p.COV_ENABLE_GUI, 0)
p.configureDebugVisualizer(p.COV_ENABLE_MOUSE_PICKING, 1)
# Zoom out camera
p.resetDebugVisualizerCamera(
cameraDistance=12.0,
cameraYaw=45,
cameraPitch=-30,
cameraTargetPosition=[0, 0, 5]
)
# Random Target
self.target_pos = np.random.uniform(self.TARGET_BOUNDS[0], self.TARGET_BOUNDS[1])
# Reset Wind (Random Walk start)
self.wind_vector = np.random.uniform(-1, 1, 3) * self.WIND_SCALE
self.wind_vector[2] = 0
# Visuals
if self.render_mode == "human":
self._draw_bounds()
self._draw_target()
self._init_wind_field()
return self._get_obs(), {}
def step(self, action):
self.step_count += 1
# 1. Update Wind (Chaotic)
self.wind_vector += np.random.normal(0, 0.1, 3)
self.wind_vector = np.clip(self.wind_vector, -5, 5)
# 2. Update Target (Moving)
# Random walk for target
target_move = np.random.normal(0, 0.05, 3)
self.target_pos += target_move
self.target_pos = np.clip(self.target_pos, self.TARGET_BOUNDS[0], self.TARGET_BOUNDS[1])
# 3. Apply Wind Force
drone = self.aviary.drones[0]
p.applyExternalForce(
drone.Id,
-1, # Base link
forceObj=self.wind_vector,
posObj=[0, 0, 0],
flags=p.LINK_FRAME
)
# 4. Step Physics
motor_command = (action + 1.0) / 2.0
self.aviary.set_all_setpoints(motor_command.reshape(1, -1))
self.aviary.step()
# 5. Get State
state = self.aviary.state(0)
pos = state[-1]
orn = state[-3]
lin_vel = state[-2]
ang_vel = state[-4]
# 6. Compute Reward
reward = self._compute_reward(pos, orn, lin_vel, ang_vel)
# 7. Check Termination
terminated = False
# Out of bounds
if (pos < self.BOUNDS[0]).any() or (pos > self.BOUNDS[1]).any():
terminated = True
reward -= 100.0 # Massive penalty for leaving
# Crash (ground)
if pos[2] < 0.1:
terminated = True
reward -= 100.0
# Target Reached
dist = np.linalg.norm(pos - self.target_pos)
if dist < 0.5:
terminated = True
reward += 100.0
truncated = self.step_count >= self.MAX_STEPS
# 8. Update Visuals
if self.render_mode == "human":
self._update_wind_field()
self._draw_target() # Update target position
# Flight Path
self.path_points.append(pos)
if len(self.path_points) > 1:
p.addUserDebugLine(self.path_points[-2], self.path_points[-1], [0, 1, 0], 2, 0)
# Info Overlay
# Clear old text
for uid in self.text_ids:
p.removeUserDebugItem(uid)
self.text_ids = []
wind_speed = np.linalg.norm(self.wind_vector)
info_text = f"Wind: {wind_speed:.2f} m/s\nDrone: {pos.round(2)}\nTarget: {self.target_pos.round(2)}"
# Draw text near the drone or fixed on screen (PyBullet text is 3D mostly, use text3d)
# Or use addUserDebugText with textPosition
uid = p.addUserDebugText(info_text, [pos[0], pos[1], pos[2] + 0.5], [0, 0, 0], textSize=1.5)
self.text_ids.append(uid)
return self._get_obs(), reward, terminated, truncated, {}
def _get_obs(self):
state = self.aviary.state(0)
pos = state[-1]
orn = state[-3]
lin_vel = state[-2]
ang_vel = state[-4]
# Relative position to target
rel_pos = self.target_pos - pos
# [rel_x, rel_y, rel_z, roll, pitch, yaw, u, v, w, p, q, r]
obs = np.concatenate([rel_pos, orn, lin_vel, ang_vel])
return obs.astype(np.float32)
def _compute_reward(self, pos, orn, lin_vel, ang_vel):
# 1. Stability (Smoothness)
# Penalize high angular velocity (spinning/shaking)
r_smooth = -0.1 * np.linalg.norm(ang_vel)
# Penalize extreme angles (flipping)
deviation = np.linalg.norm(orn[0:2])
r_stability = -deviation
# 2. Target
dist = np.linalg.norm(pos - self.target_pos)
r_target = -dist * 0.5 # Stronger distance penalty
# 3. Boundary Safety (Proximity Penalty)
# Calculate distance to nearest wall
d_min = np.min([
pos[0] - self.BOUNDS[0][0],
self.BOUNDS[1][0] - pos[0],
pos[1] - self.BOUNDS[0][1],
self.BOUNDS[1][1] - pos[1],
pos[2] - self.BOUNDS[0][2],
self.BOUNDS[1][2] - pos[2]
])
# Exponential penalty as it gets closer than 1.0m
r_boundary = 0.0
if d_min < 1.0:
r_boundary = -np.exp(1.0 - d_min) # -1 at 1m, -e at 0m
return r_stability + r_smooth + r_target + r_boundary
def _init_wind_field(self):
# Clear old arrows
for uid, _ in self.wind_arrow_ids:
p.removeUserDebugItem(uid)
self.wind_arrow_ids = []
# Create grid of arrows
# Sparse grid: 3x3x3
xs = np.linspace(self.BOUNDS[0][0], self.BOUNDS[1][0], 4)
ys = np.linspace(self.BOUNDS[0][1], self.BOUNDS[1][1], 4)
zs = np.linspace(self.BOUNDS[0][2] + 1, self.BOUNDS[1][2] - 1, 3)
for x in xs:
for y in ys:
for z in zs:
start_pos = [x, y, z]
# Initial draw (will be updated)
# Use a placeholder end_pos
end_pos = [x+0.1, y, z]
uid = p.addUserDebugLine(start_pos, end_pos, [1, 1, 0], 2)
self.wind_arrow_ids.append((uid, start_pos))
def _update_wind_field(self):
# Update all arrows to point in wind direction
# Scale wind for visual length
scale = 0.5
wind_end_offset = self.wind_vector * scale
for uid, start_pos in self.wind_arrow_ids:
end_pos = [
start_pos[0] + wind_end_offset[0],
start_pos[1] + wind_end_offset[1],
start_pos[2] + wind_end_offset[2]
]
# Color based on intensity?
intensity = np.linalg.norm(self.wind_vector)
# Map 0-5 to Color (Green to Red)
t = np.clip(intensity / 5.0, 0, 1)
color = [t, 1.0 - t, 0]
p.addUserDebugLine(
start_pos,
end_pos,
lineColorRGB=color,
lineWidth=2,
replaceItemUniqueId=uid
)
def _draw_bounds(self):
min_x, min_y, min_z = self.BOUNDS[0]
max_x, max_y, max_z = self.BOUNDS[1]
corners = [
[min_x, min_y, min_z], [max_x, min_y, min_z],
[min_x, max_y, min_z], [max_x, max_y, min_z],
[min_x, min_y, max_z], [max_x, min_y, max_z],
[min_x, max_y, max_z], [max_x, max_y, max_z]
]
lines = [
(0, 1), (1, 3), (3, 2), (2, 0), # Bottom
(4, 5), (5, 7), (7, 6), (6, 4), # Top
(0, 4), (1, 5), (2, 6), (3, 7) # Vertical
]
for start, end in lines:
p.addUserDebugLine(corners[start], corners[end], [1, 0, 0], 4)
def _draw_target(self):
if self.target_visual_id is not None:
p.removeBody(self.target_visual_id)
visual_shape = p.createVisualShape(shapeType=p.GEOM_SPHERE, radius=0.3, rgbaColor=[1, 0, 1, 0.7])
self.target_visual_id = p.createMultiBody(baseVisualShapeIndex=visual_shape, basePosition=self.target_pos)
def close(self):
self.aviary.disconnect()