_target_: pfp.policy.fm_policy.FMPolicy
  x_dim: ${x_dim}
  y_dim: ${y_dim}
  n_obs_steps: ${n_obs_steps}
  n_pred_steps: ${n_pred_steps}
  num_k_infer: 10
  time_conditioning: True
  norm_pcd_center: [0.4, 0.0, 1.4]
  augment_data: False
  noise_type: gaussian  # gaussian | trajectory | igso3
  noise_scale: 1.0
  loss_type: l2  # l2 | l1
  flow_schedule: exp  # linear | cosine | exp
  exp_scale: 4.0
  snr_sampler: uniform  # uniform | logit_normal

  obs_encoder: ${backbone}
  
  diffusion_net:
    _target_: diffusion_policy.model.diffusion.conditional_unet1d.ConditionalUnet1D
    input_dim: ${y_dim}
    global_cond_dim: "${eval: '${x_dim} * ${n_obs_steps}'}"
    diffusion_step_embed_dim: "${eval: '256 if ${model.time_conditioning} else 0'}"
    down_dims: [256, 512, 1024]
    kernel_size: 5
    n_groups: 8
    cond_predict_scale: True
    use_dropout: False

  loss_weights:
    xyz: 10.0
    rot6d: 10.0
    grip: 1.0