Skip to content

Latest commit

 

History

History
194 lines (164 loc) · 3.45 KB

EXPERIMENT_SETTING.md

File metadata and controls

194 lines (164 loc) · 3.45 KB
Raw

Experiments Settings (chronological order)

Experiments results and trained models are in the release section

F1 racetracks experiment (run-f1-tracks.zip)

Simulation, PC with i7, maps: f1_aut, f1_esp, f1_gbr, f1_mco

Use the repo as is

Testing f1tenth-rl/rl_car_driver.py

gpu-time = 0.0001
train-epoch-steps = 0
max-step-limit = 10000000
save-model-freq = 1000000

Sim2real experiment 2: transfer learning (run-unibo-roof.zip)

Training: simulation, PC with i7, unibo-roof map

Testing: real car with Jetson NX

f1tenth-rl/rl_car_driver.py

epsilon-decay = 0.99993
gpu-time = 0.018
add-velocity = False

f1tenth-rl/car/safety_control.py

USE_TTC_SIM = True
TTC_THRESHOLD_SIM = 1.21
ONLY_EXTERNAL_BARRIER = True

f1tenth-rl/car/car_control.py

MAX_SPEED_REDUCTION_SIM = 4.5
STEERING_SPEED_REDUCTION_SIM = 4.5
BACKWARD_SPEED_REDUCTION_SIM = 4.5

f1tenth-rl/car_env.py

USE_VELOCITY_AS_REWARD = False
ADD_LIDAR_DISTANCE_REWARD = False
...
self.action_set = [0, 1, 2]

Testing f1tenth-rl/rl_car_driver.py

gpu-time = 0.001
train-epoch-steps = 0
save-model-freq = 1000000
max-step-limit = 10000000
reduce-lidar-data = 36
cut-lidar-data = 5

NNs comparison experiment

Simulation, PC with i7, hairpin-track map

f1tenth-rl/rl_car_driver.py

add-velocity = False

f1tenth-rl/car/car_control.py

MAX_SPEED_REDUCTION_SIM = 3
STEERING_SPEED_REDUCTION_SIM = 3
BACKWARD_SPEED_REDUCTION_SIM = 3
BACKWARD_SECONDS_SIM = 1.8

f1tenth-rl/car/safety_control.py

EUCLIDEAN_THRESHOLD_SIM = 0.48
USE_TTC_SIM = False

f1tenth-rl/car_env.py

USE_VELOCITY_AS_REWARD = False
ADD_LIDAR_DISTANCE_REWARD = False
...
self.action_set = [0, 1, 2]
...

self.control.forward()
            reward = 0.2
...
self.control.right()
            reward = 0.05
...
self.control.left()
            reward = 0.05

CNN1D (run-hairpin-track-simulator-cnn1d.zip)

f1tenth-rl/rl_car_driver.py

epsilon-decay = 0.9999342
observation-steps = 400
gpu-time = 0.014
reduce-lidar-data = 30
cut-lidar-data = 8

Dense (run-hairpin-track-simulator-dense.zip)

same as CNN1D, plus:

f1tenth-rl/dqn.py

    def __build_q_net(self):
        ...
            return self.__build_dense()

CNN2D (run-hairpin-track-simulator-cnn2d.zip)

same as CNN1D, plus:

f1tenth-rl/rl_car_driver.py

epsilon-decay = 0.99992
slowdown-cycle = False
lidar-to-image = True

Sim2real experiment 1: training on the physical car (run-real-car.zip)

Jetson TX2 - slow speed (1/8 of max speed)

f1tenth-rl/car/car_control.py

MAX_SPEED_REDUCTION = 8
STEERING_SPEED_REDUCTION = 8
BACKWARD_SPEED_REDUCTION = 8
BACKWARD_SECONDS = 1.8

f1tenth-rl/car/safety_control.py

USE_TTC = True
TTC_THRESHOLD_REAL_CAR = 0.62
EUCLIDEAN_THRESHOLD_REAL_CAR = 0.08

f1tenth-rl/rl_car_driver.py

epsilon-decay = 0.9998
observation-steps = 500
repeat-action = 5
gpu-time = 0.03
reduce-lidar-data = 36
cut-lidar-data = 5
max-distance-norm = 5
save-model-freq = 5000
add-velocity = False

f1tenth-rl/car_env.py

USE_VELOCITY_AS_REWARD = True
ADD_LIDAR_DISTANCE_REWARD = True
LIDAR_DISTANCE_WEIGHT = 0.1
VELOCITY_NORMALIZATION = 0.55
...
self.action_set = [0, 1, 2]

Evaluation only - high speed (1/4 of max speed)

Same as above, plus:

f1tenth-rl/rl_car_driver.py

gpu-time = 0.001
train-epoch-steps = 0
save-model-freq = 1000000
max-step-limit = 10000000

f1tenth-rl/car/car_control.py

MAX_SPEED_REDUCTION = 4
STEERING_SPEED_REDUCTION = 4
BACKWARD_SECONDS = 2