All configuration defaults
Generated from the typed configuration. See configuration precedence.
Explicit presets
Presets apply before project YAML and CLI overrides.
smoke
train:
name: smoke
preset: smoke
num_envs: 2
max_iterations: 2
save_interval: 1
validation_interval: 0
data:
train_maps: 1
validation_maps: 1
test_maps: 1
pairs_per_map: 3
navigation
train:
name: navigation
preset: navigation
validation_interval: 200
Defaults
schema: dwarl/1
robot: null
train:
preset: null
name: navigation
num_envs: 64
seed: 42
device: auto
max_iterations: 3000
resume: null
show: false
detached: false
save_interval: 50
validation_interval: 0
validation_cases: 32
benchmark_interval: 0
benchmark_cases: 32
early_stop:
enabled: false
patience: 5
min_delta: 0.01
deployment_ready: true
data:
path: null
allow_custom: false
train_maps: 128
validation_maps: 16
test_maps: 16
pairs_per_map: 32
workers: 1
test:
run: latest
checkpoint: latest
suite: benchmark
sim: mujoco
case: null
show: false
repeats: 1
seed: 42
dataset: null
max_cases: 32
controllers:
- dwarl
profile_compute: false
controller_params: null
startup_timeout: 120.0
stall_timeout: 30.0
deploy:
run: latest
checkpoint: latest
recipe:
actions:
drive:
class_type: dwarl.navigation.actions:DifferentialDriveAction
params:
mode: window
clip:
- -1.0
- 1.0
commands:
path:
class_type: dwarl.navigation.commands:PathCommand
params:
goal_mode: lookahead
lookahead_distances:
- 0.5
- 1.0
- 2.0
- 3.0
lookahead_distance: 2.0
projection_window:
- -5
- 30
observations:
lidar:
concatenate_dim: -1
terms:
scan:
func: dwarl.navigation.observations:LidarHistory
params:
history:
- 0
- 2
- 5
normalization: log
clip: null
scale: 1.0
state:
concatenate_dim: -1
terms:
goals:
func: dwarl.navigation.observations:goal_points
params:
command_name: path
max_distance: 10.0
clip: null
scale: 1.0
last_action:
func: dwarl.navigation.observations:last_action
params: {}
clip: null
scale: 1.0
velocity:
func: dwarl.navigation.observations:base_velocity
params: {}
clip: null
scale: 1.0
command_feedback: null
privileged:
concatenate_dim: -1
terms:
path:
func: dwarl.navigation.observations:path_features
params:
command_name: path
lookahead:
- 0.5
- 1.0
- 2.0
- 3.0
clip: null
scale: 1.0
rewards:
progress:
func: dwarl.navigation.rewards:progress
params: {}
weight: 10.0
cross_track:
func: dwarl.navigation.rewards:cross_track
params: {}
weight: 2.0
heading:
func: dwarl.navigation.rewards:heading
params: {}
weight: 1.0
goal:
func: dwarl.navigation.rewards:goal
params: {}
weight: 300.0
collision:
func: dwarl.navigation.rewards:collision
params: {}
weight: 500.0
stuck:
func: dwarl.navigation.rewards:stuck
params: {}
weight: 200.0
time:
func: dwarl.navigation.rewards:time
params: {}
weight: 0.5
action_rate:
func: dwarl.navigation.rewards:action_rate
params: {}
weight: 0.2
proximity:
func: dwarl.navigation.rewards:proximity
params:
band: 0.25
power: 1.0
weight: 5.0
reverse:
func: dwarl.navigation.rewards:reverse
params:
deadband: 0.02
power: 1.0
weight: 8.0
terminations:
collision:
func: dwarl.navigation.terminations:collision
params:
min_upright: 0.5
time_out: false
role: collision
flipped:
func: dwarl.navigation.terminations:flipped
params:
min_upright: 0.5
time_out: false
role: collision
success:
func: dwarl.navigation.terminations:goal_reached
params:
command_name: path
tolerance: 0.3
min_upright: 0.5
time_out: false
role: success
stuck:
func: dwarl.navigation.terminations:NoProgress
params:
command_name: path
steps: 50
distance: 0.3
time_out: false
role: stuck
timeout:
func: dwarl.navigation.terminations:PathTimeLimit
params:
command_name: path
min_steps: 100
time_factor: 2.5
time_out: true
role: null
sim:
timestep: 0.005
tile_gap: 1.0
pool_size: 300
num_shards: 0
render_fps: 30
realtime: false
realtime_speed: 1.0
env:
control_dt: 0.1
episodes_per_map: 1
spawn_yaw_noise: 3.14159
spawn_yaw_curriculum: true
spawn_yaw_min: 0.5
spawn_yaw_full_level: 0.6
curriculum:
enabled: true
init_level: 0.2
min_level: 0.1
step: 0.05
window: 200
promote_success: 0.75
demote_success: 0.3
agent:
num_steps_per_env: 64
logger: tensorboard
obs_groups:
actor:
- lidar
- state
critic:
- lidar
- state
- privileged
actor:
class_name: dwarl.training.models.lidar_cnn:LidarCNNModel
hidden_dims:
- 512
- 256
- 128
activation: elu
obs_normalization: true
encoder_cfg:
channels:
- 32
- 64
- 64
kernels:
- 7
- 5
- 3
strides:
- 2
- 2
- 2
latent_dim: 128
activation: elu
distribution_cfg:
class_name: GaussianDistribution
init_std: 0.6
std_type: scalar
critic:
class_name: dwarl.training.models.lidar_cnn:LidarCNNModel
hidden_dims:
- 512
- 256
- 128
activation: elu
obs_normalization: true
encoder_cfg:
channels:
- 32
- 64
- 64
kernels:
- 7
- 5
- 3
strides:
- 2
- 2
- 2
latent_dim: 128
activation: elu
distribution_cfg: null
algorithm:
class_name: PPO
num_learning_epochs: 5
num_mini_batches: 4
clip_param: 0.2
gamma: 0.99
lam: 0.95
value_loss_coef: 1.0
entropy_coef: 0.005
learning_rate: 0.0003
max_grad_norm: 1.0
schedule: adaptive
desired_kl: 0.01