evaluation.ini

; =============================================================================
; Evaluation Framework Configuration
; =============================================================================
; Override any parameter via CLI using dot-notation:
;   python pufferlib/ocean/benchmark/eval.py --planner.type pdm --traffic.type ppo --eval.viz True
;
; Sections:
;   [env]           - Drive C environment settings (parsed by C ini handler)
;   [eval]          - General evaluation settings
;   [planner]       - Ego planner selection (type = pdm|ppo|idm|smart|hybrid|constant_velocity)
;   [planner.<type>]- Per-planner defaults
;   [traffic]       - Traffic agent controller selection
;   [traffic.<type>]- Per-controller defaults
; =============================================================================

; Drive C environment defaults for evaluation
[env]
action_type = continuous
episode_length = 91
goal_behavior = 3
termination_mode = 1
collision_behavior = 2
offroad_behavior = 2
dt = 0.1
collision_shrink = 1.0
reward_vehicle_collision = -0.5
reward_offroad_collision = -0.5
; Goal reward — without this, env->reward_goal defaults to 0 and reaching
; the goal yields no reward contribution. Matches training (Gigaflow paper).
reward_goal = 1.0
reward_goal_post_respawn = 0.25

[eval]
split = pufferinter
episode_length = 91
action_type = continuous
; Drive dynamics model: "classic" (accel + steer) or "jerk" (jerk integration in C).
; Required value for DriveConditionedPaper checkpoints is "jerk".
dynamics_model = classic
; Reward conditioning: when 1, env appends 9 per-agent creward coefficients to obs
; AND uses them in reward computation. Required value for DriveConditionedPaper is 1.
reward_conditioning = 0
; Scene visualization (road, agents, per-step PNGs -> GIF)
viz = False
; Planner-specific visualization (PDM proposals, etc.)
planner_viz = False
goal_behavior = 3
; Probability per goal-reach that the new goal is sampled on a parallel
; lane instead of the current one (only active when goal_behavior=5).
goal_lane_change_prob = 0.0
; Arc-length ahead of the agent at which the new goal is placed (m).
goal_target_distance = 20.0
termination_mode = 1
collision_behavior = 2
offroad_behavior = 2
; Uncertainty estimation (requires ensemble weights)
compute_uncertainty = False
ensemble_weight_paths =


; =============================================================================
; Ego Planner
; =============================================================================
[planner]
type = pdm

[planner.pdm]
horizon = 40
proposal_other = constant_velocity
velocity_fractions = (0.2, 0.4, 0.6, 0.8, 1.0)
lateral_offsets = (-1.0, 0.0, 1.0)
; Maximum velocity in m/s for IDM proposals (fractions are multiplied by this)
max_velocity = 25.0
min_gap = 1.0
headway_time = 1.5
accel_max = 1.5
decel_max = 3.0

[planner.ppo]
weights_path =
device = cuda
ensemble_weights =

[planner.idm]
target_velocity = 15.0
min_gap = 1.0
headway_time = 1.5
accel_max = 1.0
decel_max = 3.0

[planner.smart]
weights_path =
device = cuda
temperature = 1.0
greedy = True
repredict_interval = 5

[planner.hybrid]
ppo_weights_path =
ensemble_weights =
epistemic_threshold = 0.8
pdm_min_steps = 1
force_ppo = false
force_pdm = false
switch_mode = pdm_score
value_variance_threshold = 0.1
lookahead_steps = 0
ppo_mode = single_step
ppo_rollout_strategy = beam_search
ppo_rollout_top_k = 8
ppo_rollout_beam_width = 4
ppo_rollout_branch_factor = 4
ppo_rollout_horizon = 10
ppo_rollout_w_cmf = 0.333
ppo_rollout_w_align = 0.333
ppo_rollout_w_ctr = 0.333
ppo_rollout_lane_dist_scale = 2.0
device = cuda

[planner.constant_velocity]

[planner.conditioned_mix]
weights_path =
device = cuda
creward.delta_goal = 0.0
creward.alpha_collision = 3.0
creward.alpha_boundary = 3.0
creward.alpha_comfort = 0.05
creward.alpha_l_align = 0.015
creward.alpha_vel_align = 0.5
creward.alpha_l_center = 0.0038
creward.alpha_center_bias = 0.0
creward.alpha_reverse = 0.0038
creward.goal_speed = 20.0

[planner.conditioned_aggr]
weights_path =
device = cuda
creward.delta_goal = 0.0
creward.alpha_collision = 0.2
creward.alpha_boundary = 0.2
creward.alpha_comfort = 0.0
creward.alpha_l_align = 0.001
creward.alpha_vel_align = 0.1
creward.alpha_l_center = 0.00025
creward.alpha_center_bias = 0.0
creward.alpha_reverse = 0.0005
creward.goal_speed = 30.0

[planner.conditioned_normal]
weights_path =
device = cuda
creward.delta_goal = 0.0
creward.alpha_collision = 3.0
creward.alpha_boundary = 3.0
creward.alpha_comfort = 0.05
creward.alpha_l_align = 0.015
creward.alpha_vel_align = 0.5
creward.alpha_l_center = 0.0038
creward.alpha_center_bias = 0.0
creward.alpha_reverse = 0.0038
creward.goal_speed = 20.0

[planner.conditioned_caut]
weights_path =
device = cuda
creward.delta_goal = 0.0
creward.alpha_collision = 3.0
creward.alpha_boundary = 3.0
creward.alpha_comfort = 0.1
creward.alpha_l_align = 0.025
creward.alpha_vel_align = 1.0
creward.alpha_l_center = 0.0075
creward.alpha_center_bias = 0.5
creward.alpha_reverse = 0.0075
creward.goal_speed = 5.0

; =============================================================================
; Traffic Agent Controller
; =============================================================================
[traffic]
type = idm

[traffic.pdm]
horizon = 40
proposal_other = constant_velocity

[traffic.ppo]
weights_path =
device = cuda

[traffic.idm]
target_velocity = 15.0
min_gap = 1.0
headway_time = 1.5
accel_max = 1.0
decel_max = 3.0

[traffic.smart]
weights_path =
device = cuda
temperature = 1.0
greedy = True
repredict_interval = 5

[traffic.expert]

[traffic.constant_velocity]

[traffic.conditioned_mix]
weights_path =
device = cuda
creward_profiles = [
    (0.2, 0.2, 0.0,  0.001, 0.1, 0.00025, 0.0, 0.0005, 30.0),
    (3.0, 3.0, 0.05, 0.015, 0.5, 0.0038,  0.0, 0.0038, 20.0),
    (3.0, 3.0, 0.1,  0.025, 1.0, 0.0075,  0.5, 0.0075,  5.0),
    ]

[traffic.conditioned_aggr]
weights_path =
device = cuda
creward_profiles = [
    (0.2, 0.2, 0.0,  0.001, 0.1, 0.00025, 0.0, 0.0005, 30.0),
    ]

[traffic.conditioned_normal]
weights_path =
device = cuda
creward_profiles = [
    (3.0, 3.0, 0.05, 0.015, 0.5, 0.0038,  0.0, 0.0038, 20.0),
    ]

[traffic.conditioned_caut]
weights_path =
device = cuda
creward_profiles = [
    (3.0, 3.0, 0.1,  0.025, 1.0, 0.0075,  0.5, 0.0075,  5.0),
    ]