Skip to content

rldriving: curvature-rate cost and lower longitudinal smoothing #80

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
haraschax merged 10 commits into from
Sep 16, 2026
Merged

rldriving: curvature-rate cost and lower longitudinal smoothing #80

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
2dde3c0
rldriving: use curvature rate and annealed acceleration exploration
haraschax Sep 11, 2026
223a4ee
rldriving: simplify rollout exploration
haraschax Sep 11, 2026
6eeaf3f
rldriving: simplify exploration scheduling
haraschax Sep 11, 2026
756c189
rldriving: default exploration to 0.5
haraschax Sep 11, 2026
7914313
rldriving: remove curvature cost
haraschax Sep 11, 2026
bcb21ff
rldriving: simplify curvature-rate calculation
haraschax Sep 11, 2026
8469fa0
rldriving: remove duplicate configuration forwarding
haraschax Sep 11, 2026
2cb59d7
rldriving: remove trainer exploration configuration
haraschax Sep 11, 2026
c125d62
rldriving: restore reference warmup and final learning rate
haraschax Sep 14, 2026
afa31b3
rldriving: lower longitudinal smoothing to 0.05
haraschax Sep 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions torchtitan/experiments/rldriving/config_registry.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,12 +96,12 @@ def rldriving() -> RLDrivingTrainer.Config:
gamma=0.95,
fps=fps,
smooth_lat_cost=0.15,
smooth_long_cost=0.1,
curv_cost=100.0,
smooth_long_cost=0.05,
curv_rate_cost=20.0,
),
warm_start_checkpoint=os.getenv(
"RLDRIVING_WARM_START_CHECKPOINT",
"849a624a-8a7d-8946-bf04-86148e5e0ef8/56320",
"b9facbcc-4d47-410e-b3ce-dfcbad12ba92/56320",
),
tokenizer=NoOpTokenizer.Config(),
dataloader=RLDrivingDataLoader.Config(
Expand Down
53 changes: 21 additions & 32 deletions torchtitan/experiments/rldriving/loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,14 +38,14 @@ def _sample_fixed_noise_policy(

def _critic_loss(
*,
config: RLDrivingLoss.Config,
bootstrap_actor_outputs: ActorOutputs,
targets: Targets,
online_critic: nn.Module,
target_critic: nn.Module,
current_inputs: ModelInputs,
bootstrap_inputs: ModelInputs,
action_noise_A: torch.Tensor,
gamma: float,
) -> LossResult:
action_reward_B = targets["action_reward"]
rollout_action_BA = action_reward_B[:, 0:2]
Expand All @@ -66,11 +66,11 @@ def _critic_loss(
action=bootstrap_action_BA,
)
bootstrap_B = torch.minimum(q1_target_B, q2_target_B)
discounts_N = gamma ** torch.arange(
discounts_N = config.gamma ** torch.arange(
rewards_BN.shape[1], device=rewards_BN.device, dtype=rewards_BN.dtype
)
discounted_reward_B = (rewards_BN * discounts_N).sum(dim=1)
bootstrap_discount = gamma ** rewards_BN.shape[1]
bootstrap_discount = config.gamma ** rewards_BN.shape[1]
target_B = discounted_reward_B + bootstrap_discount * bootstrap_B
q_target_abs_gap_B = torch.abs(q1_target_B - q2_target_B)
q_rollout_abs_gap_B = torch.abs(q1_rollout_B - q2_rollout_B)
Expand All @@ -94,19 +94,15 @@ def _critic_loss(

def _actor_loss(
*,
config: RLDrivingLoss.Config,
actor_outputs: ActorOutputs,
next_actor_outputs: ActorOutputs,
online_critic: nn.Module,
current_inputs: ModelInputs,
targets: Targets,
fps: float,
smooth_lat_cost: float,
smooth_long_cost: float,
curv_cost: float,
action_bound: float,
action_bound_loss_weight: float,
) -> LossResult:
action_pred_BA = actor_outputs[ACTION_OUTPUT]
next_action_pred_BA = next_actor_outputs[ACTION_OUTPUT]
q1_new_B, q2_new_B = online_critic(
inputs=current_inputs,
action=action_pred_BA[:, :2],
Expand All @@ -115,26 +111,30 @@ def _actor_loss(
actor_q_abs_gap_B = torch.abs(q1_new_B - q2_new_B)

curvature_B = action_pred_BA[:, 0] / targets["speed"].squeeze(-1).square()
curvature_loss_B = curv_cost * curvature_B.square()
next_curvature_B = next_action_pred_BA[:, 0] / targets["next_speed"].squeeze(-1).square()
curvature_rate_B = (next_curvature_B - curvature_B) * config.fps
curvature_rate_loss_B = config.curv_rate_cost * curvature_rate_B.square()

command_jerk_BA = (next_actor_outputs[ACTION_OUTPUT][:, :2] - action_pred_BA[:, :2]).abs() * fps
smooth_lat_B = smooth_lat_cost * command_jerk_BA[:, 0].square()
smooth_long_B = smooth_long_cost * command_jerk_BA[:, 1].square()
command_jerk_BA = (next_action_pred_BA[:, :2] - action_pred_BA[:, :2]).abs() * config.fps
smooth_lat_B = config.smooth_lat_cost * command_jerk_BA[:, 0].square()
smooth_long_B = config.smooth_long_cost * command_jerk_BA[:, 1].square()
smooth_B = smooth_lat_B + smooth_long_B
actor_loss_B = actor_pi_B + curvature_loss_B + smooth_B
actor_loss_B = actor_pi_B + curvature_rate_loss_B + smooth_B

action_abs_BA = torch.abs(action_pred_BA[..., :2])
action_bound_excess_BA = torch.clamp(action_abs_BA - action_bound, min=0.0)
action_bound_excess_BA = torch.clamp(action_abs_BA - config.action_bound, min=0.0)
action_bound_loss_B = action_bound_excess_BA.square().mean(dim=-1)
loss_B = actor_loss_B + action_bound_loss_weight * action_bound_loss_B
loss_B = actor_loss_B + config.action_bound_loss_weight * action_bound_loss_B

metrics = {
"loss": loss_B.detach(),
"actor_loss": actor_loss_B.detach(),
"actor_pi": actor_pi_B.detach(),
"actor_q_abs_gap": actor_q_abs_gap_B.detach(),
"actor_curv": curvature_B.detach(),
"actor_curv_loss": curvature_loss_B.detach(),
"actor_curv_rate": curvature_rate_B.detach(),
"actor_curv_rate_abs": curvature_rate_B.abs().detach(),
"actor_curv_rate_loss": curvature_rate_loss_B.detach(),
"actor_cmd_lat_jerk": command_jerk_BA[:, 0].detach(),
"actor_cmd_long_jerk": command_jerk_BA[:, 1].detach(),
"actor_smooth_lat_loss": smooth_lat_B.detach(),
Expand All @@ -155,7 +155,7 @@ class Config(BaseLoss.Config):
fps: float
smooth_lat_cost: float = 0.0
smooth_long_cost: float = 0.0
curv_cost: float = 0.0
curv_rate_cost: float = 0.0
action_bound: float = 10.0
action_bound_loss_weight: float = 1.0

Expand All @@ -165,14 +165,8 @@ def __init__(
*,
compile_config: CompileConfig | None = None,
) -> None:
self.config = config
self.action_noise_A = torch.tensor(config.action_noise)
self.gamma = config.gamma
self.fps = config.fps
self.smooth_lat_cost = config.smooth_lat_cost
self.smooth_long_cost = config.smooth_long_cost
self.curv_cost = config.curv_cost
self.action_bound = config.action_bound
self.action_bound_loss_weight = config.action_bound_loss_weight

self.critic_fn = _critic_loss
self.actor_fn = _actor_loss
Expand Down Expand Up @@ -204,14 +198,14 @@ def critic_loss(
bootstrap_inputs: ModelInputs,
) -> LossResult:
return self.critic_fn(
config=self.config,
bootstrap_actor_outputs=bootstrap_actor_outputs,
targets=targets,
online_critic=online_critic,
target_critic=target_critic,
current_inputs=current_inputs,
bootstrap_inputs=bootstrap_inputs,
action_noise_A=self.action_noise_A,
gamma=self.gamma,
)

def actor_loss(
Expand All @@ -224,15 +218,10 @@ def actor_loss(
targets: Targets,
) -> LossResult:
return self.actor_fn(
config=self.config,
actor_outputs=actor_outputs,
next_actor_outputs=next_actor_outputs,
online_critic=online_critic,
current_inputs=current_inputs,
targets=targets,
fps=self.fps,
smooth_lat_cost=self.smooth_lat_cost,
smooth_long_cost=self.smooth_long_cost,
curv_cost=self.curv_cost,
action_bound=self.action_bound,
action_bound_loss_weight=self.action_bound_loss_weight,
)