feat(policy): add minimal swing policy for hil record smoke runs

这个提交包含在:
Elvin
2026-03-07 12:45:43 +08:00
父节点 70c4b2d665
当前提交 0714316bab
共有 5 个文件被更改,包括 159 次插入 和 0 次删除

查看文件

@ -20,6 +20,7 @@ from .pi0_fast.configuration_pi0_fast import PI0FastConfig as PI0FastConfig
from .pi05.configuration_pi05 import PI05Config as PI05Config
from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
from .smolvla.processor_smolvla import SmolVLANewLineProcessor
from .swing.configuration_swing import SwingConfig as SwingConfig
from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
from .wall_x.configuration_wall_x import WallXConfig as WallXConfig
@ -32,6 +33,7 @@ __all__ = [
"PI05Config",
"PI0FastConfig",
"SmolVLAConfig",
"SwingConfig",
"SARMConfig",
"TDMPCConfig",
"VQBeTConfig",

查看文件

@ -0,0 +1,6 @@
#!/usr/bin/env python
from .configuration_swing import SwingConfig as SwingConfig
from .modeling_swing import SwingPolicy as SwingPolicy
__all__ = ["SwingConfig", "SwingPolicy"]

查看文件

@ -0,0 +1,48 @@
#!/usr/bin/env python
from dataclasses import dataclass
from lerobot.configs.policies import PreTrainedConfig
from lerobot.optim.optimizers import AdamConfig
@PreTrainedConfig.register_subclass("swing")
@dataclass
class SwingConfig(PreTrainedConfig):
n_obs_steps: int = 1
swing_joint_index: int = 0
swing_amplitude: float = 0.08
swing_period_steps: int = 60
center_from_observation: bool = True
optimizer_lr: float = 1e-4
def __post_init__(self) -> None:
super().__post_init__()
if self.swing_period_steps <= 0:
raise ValueError("`swing_period_steps` must be > 0.")
if self.swing_amplitude < 0:
raise ValueError("`swing_amplitude` must be >= 0.")
def get_optimizer_preset(self) -> AdamConfig:
return AdamConfig(lr=self.optimizer_lr)
def get_scheduler_preset(self) -> None:
return None
def validate_features(self) -> None:
if self.action_feature is None:
raise ValueError("Swing policy requires an action feature.")
if not self.action_feature.shape:
raise ValueError("Swing policy action feature shape must be non-empty.")
@property
def observation_delta_indices(self) -> None:
return None
@property
def action_delta_indices(self) -> None:
return None
@property
def reward_delta_indices(self) -> None:
return None

查看文件

@ -0,0 +1,60 @@
#!/usr/bin/env python
import math
from typing import Any
import torch
from torch import Tensor, nn
from lerobot.policies.pretrained import ActionSelectKwargs, PreTrainedPolicy
from lerobot.utils.constants import OBS_STATE
from .configuration_swing import SwingConfig
class SwingPolicy(PreTrainedPolicy):
config_class = SwingConfig
name = "swing"
def __init__(self, config: SwingConfig, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self._step = 0
self._anchor = nn.Parameter(torch.zeros(1), requires_grad=True)
def get_optim_params(self) -> dict:
return {"params": [self._anchor]}
def reset(self):
self._step = 0
def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict | None]:
loss = self._anchor.sum() * 0.0
return loss, None
def _action_dim(self) -> int:
action_feature = self.config.action_feature
if action_feature is None or not action_feature.shape:
raise ValueError("Swing policy requires a valid action feature shape.")
return int(action_feature.shape[0])
def predict_action_chunk(self, batch: dict[str, Tensor], **kwargs: ActionSelectKwargs) -> Tensor:
return self.select_action(batch, **kwargs)
def select_action(self, batch: dict[str, Tensor], **kwargs: ActionSelectKwargs) -> Tensor:
action_dim = self._action_dim()
action = torch.zeros((1, action_dim), device=self._anchor.device, dtype=torch.float32)
if self.config.center_from_observation and OBS_STATE in batch:
obs_state = batch[OBS_STATE]
if obs_state.dim() == 1:
obs_state = obs_state.unsqueeze(0)
if obs_state.shape[-1] >= action_dim:
action = obs_state[:, :action_dim].to(dtype=torch.float32).clone()
period = max(self.config.swing_period_steps, 1)
phase = 2.0 * math.pi * ((self._step % period) / period)
offset = self.config.swing_amplitude if math.sin(phase) >= 0 else -self.config.swing_amplitude
joint_idx = max(0, min(self.config.swing_joint_index, action_dim - 1))
action[:, joint_idx] = action[:, joint_idx] + offset
self._step += 1
return action

查看文件

@ -0,0 +1,43 @@
#!/usr/bin/env python
from typing import Any
from lerobot.processor import (
AddBatchDimensionProcessorStep,
DeviceProcessorStep,
PolicyAction,
PolicyProcessorPipeline,
RenameObservationsProcessorStep,
)
from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
from .configuration_swing import SwingConfig
def make_swing_pre_post_processors(
config: SwingConfig,
dataset_stats: dict[str, dict[str, Any]] | None = None,
) -> tuple[
PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
PolicyProcessorPipeline[PolicyAction, PolicyAction],
]:
del dataset_stats
pre_steps = [
RenameObservationsProcessorStep(rename_map={}),
AddBatchDimensionProcessorStep(),
DeviceProcessorStep(device=config.device),
]
post_steps = [DeviceProcessorStep(device="cpu")]
return (
PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
steps=pre_steps,
name=POLICY_PREPROCESSOR_DEFAULT_NAME,
),
PolicyProcessorPipeline[PolicyAction, PolicyAction](
steps=post_steps,
name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
to_transition=policy_action_to_transition,
to_output=transition_to_policy_action,
),
)