code-HFD / src /psi /config /model_act.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
06f9561 verified
Raw History Blame Contribute Delete
1.04 kB
from tyro.conf import subcommand as cmd
from typing import Union, Annotated, Optional, List
from psi.config.config import ModelConfig
from pydantic import Field
class ACTModelConfig(ModelConfig):
n_obs_steps: int = 1
chunk_size: int = 100 # action chunk size (prediction horizon)
n_action_steps: int = 100 # number of action steps to execute
# Dimensions
action_dim: int = 7 # action dimension
state_dim: int = 15 # robot state dimension (proprioception)
# Transformer architecture
dim_model: int = 512
n_heads: int = 8
dim_feedforward: int = 3200
feedforward_activation: str = "relu"
n_encoder_layers: int = 4
n_decoder_layers: int = 1 # Original ACT has a bug, only 1 layer is used
pre_norm: bool = False
dropout: float = 0.1
# VAE configuration
use_vae: bool = True
latent_dim: int = 32
n_vae_encoder_layers: int = 4
kl_weight: float = 10.0
# Inference
temporal_ensemble_coeff: float | None = None # set to 0.01 to enable