"""ACT (Psi0 baselines/act = LeRobot ACT, ResNet18) launch config for the Humanoid Benchmark tasks. Registered by hb_act_train.py as psi.config.train.hb_act_config; the pinned upstream scripts/train.py loads it unchanged and Trainer.instantiate picks psi.trainers.act_g1.ActG1Trainer (--train.name=act-g1). Native recipe = baselines/act/train_act_g1_simple.sh (G1 simulation): chunk 16, VAE (kl 10, latent 32), dim 512, 4 encoder / 1 decoder layers, ResNet18 ImageNet backbone, images resized to 256x480 (NEAREST) and center-cropped to 224x224, ImageNet normalisation, bounds-normalised actions and states, AdamW lr 1e-4 (betas 0.95/0.999, wd 1e-6), cosine schedule, warmup 2.5 % (1000/40000), bf16 DDP. HB changes (CLAUDE.md v2, identical inputs for every model): * repack = hb_psi0_config.HbSonicRepackTransform (the same Psi0 view as Psi-Zero: G1 observation.hb44 / action.hb76, H2 observation.hb78 / action.hb75, cameras cam0 + chest, raw state WITHOUT joint velocity), with the HB v2 proprio augmentation (15 % delay 1..2 frames, +-2 deg per-trajectory bias, 0.25 deg noise). * field = native SonicActionStateTransform (bounds), as Psi-Zero (same stats_psi0.json of the view). * model = HbACTModelTransform: native geometry (resize 256x480 NEAREST, center crop 224x224) and normalisation; native ColorJitter (img_aug) OFF, replaced by hb_aug_v2.augment on both cameras jointly at 224x224. ACT natively encodes every camera with the shared ResNet18 and concatenates the feature maps (n_cameras = 2). """ import os import sys from typing import Any import numpy as np import torch from pydantic import model_validator from torchvision.transforms import v2 from typing_extensions import Self _HERE = os.path.dirname(os.path.abspath(__file__)) _PSI0 = os.path.join(os.path.dirname(_HERE), "hb_psi0") if _PSI0 not in sys.path: sys.path.insert(0, _PSI0) import hb_psi0_config as HP # noqa: E402 HbSonicRepackTransform, hb_rng, set_epoch, A = hb_aug_v2 from psi.config import transform as pt # noqa: E402 from psi.config import transform_psi0_sonic as ps # noqa: E402 from psi.config.config import LaunchConfig # noqa: E402 from psi.config.data_lerobot import LerobotDataConfig # noqa: E402 from psi.config.model_act import ACTModelConfig # noqa: E402 from psi.config.transform import DataTransform # noqa: E402 RAW_STATE_DIMS = (44, 78) # G1 / H2 benchmark state input (no joint velocity) class HbACTModelTransform(pt.ACTModelTransform): hb_aug: bool = True def __call__(self, data: dict[str, Any], no_aug: bool = False, **kwargs) -> dict[str, Any]: assert not self.img_aug, "ACT native ColorJitter must stay off (HB v2 replaces it)" geom = v2.Compose([v2.ToImage(), v2.ToDtype(torch.float32, scale=True), self.resize(), self.center_crop()]) x = torch.stack([geom(img) for img in data["observations"]]) # (V, 3, 224, 224) in [0, 1], native geometry if self.hb_aug and not no_aug: y, p = HP.A.augment(x[None], HP.hb_rng()) # [T=1, V, 3, H, W]: colour shared, geometry/light per camera x = y[0].clamp(0.0, 1.0) HP._log_aug(p, y.shape) images = self.normalize()(x) mask = np.asarray(data["actions_mask"]).astype(bool) if not mask.all(): raise ValueError("HB views never pad actions (actions_mask must be all True)") return { "observation.images": images, # (n_cameras, C, H, W) "observation.state": data["states"], # (1, Ds) "action": data["actions"], # (chunk, Da) "action_is_pad": np.zeros(np.asarray(data["actions"]).shape[0], dtype=bool), } class HbACTDataTransform(DataTransform): repack: HP.HbSonicRepackTransform field: ps.SonicActionStateTransform model: HbACTModelTransform class HbACTDataConfig(LerobotDataConfig): transform: HbACTDataTransform # must be re-annotated for tyro (see hb_psi0_config.HbDataConfig) num_workers: int | None = None prefetch_factor: int | None = None class DynamicLaunchConfig(LaunchConfig): data: HbACTDataConfig model: ACTModelConfig @model_validator(mode="after") def check_hb_dims(self, __context: Any) -> Self: repack, field, model = self.data.transform.repack, self.data.transform.field, self.model assert repack.pad_action_dim == field.pad_action_dim == model.action_dim, \ (repack.pad_action_dim, field.pad_action_dim, model.action_dim) assert repack.pad_state_dim == field.pad_state_dim == model.state_dim, \ (repack.pad_state_dim, field.pad_state_dim, model.state_dim) assert model.state_dim in RAW_STATE_DIMS, f"ACT state must be the raw 44 / 78 dims (no velocity), got {model.state_dim}" assert model.chunk_size == model.n_action_steps == repack.action_chunk_size, \ (model.chunk_size, model.n_action_steps, repack.action_chunk_size) assert field.action_min is not None and len(field.action_min) == model.action_dim, "action stats width" assert field.normalize_state and field.state_min is not None and len(field.state_min) == model.state_dim, \ "state stats width" assert not self.data.transform.model.img_aug, "disable --data.transform.model.img-aug (HB v2 replaces it)" assert field.state_noise_std == 0.0 and repack.state_temporal_jitter == 0 assert 0 <= repack.hb_body_dims <= model.state_dim return self