datasets: vla_data: action_mode: abs data_mix: robotwin_clean_50 data_root_dir: ./playground/Datasets/RoboTwin dataset_py: lerobot_datasets per_device_batch_size: 48 sequential_step_sampling: false video_backend: pyav framework: action_model: act_dim_feedforward: 2048 act_dropout: 0.1 act_hidden_dim: 256 act_loss_type: l1 act_mlp_hidden_dim: 512 act_nheads: 8 act_num_layers: 3 act_state_hidden_dim: 256 act_state_tokens: 2 action_dim: 14 action_hidden_dim: 1024 action_horizon: 50 action_model_type: act diffusion_model_cfg: cross_attention_dim: 256 hidden_size: 1024 state_dim: 14 dinov3: attn_implementation: flash_attention_2 freeze_vision_encoder: false image_size: 224 local_files_only: false model_path: facebook/dinov3-vitl16-pretrain-lvd1689m num_views: 3 vision_dropout: 0.1 vision_pos_init_std: 0.01 vision_pos_scale_init: 0.01 fusion: enhancer_inner_dim: 1024 fusion_dropout: 0.0 fusion_droppath: 0.1 hidden_dim: 256 nheads: 8 num_layers: 6 text_dropout: 0.0 groundingdino: load_bert: false load_feature_enhancer: false load_pretrained: false load_text_proj: false pretrained_ckpt: null name: GroundingDINODiT text: attn_implementation: flash_attention_2 bert_path: google-bert/bert-base-uncased freeze_text_encoder: true local_files_only: false max_text_len: 256 sub_sentence_present: true output_dir: . run_id: turbovla_robotwin_clean50_55k_ema run_root_dir: . seed: 42 trainer: ema_decay: 0.999 ema_device: cuda eval_interval: 1000 freeze_modules: text_encoder.bert gradient_clipping: 1.0 learning_rate: action_model: 5.0e-05 base: 5.0e-05 dinov3: 5.0e-05 feature_enhancer: 5.0e-05 text_encoder: 5.0e-05 vision_proj: 5.0e-05 logging_frequency: 50 lr_scheduler_type: warmup_constant_with_factor max_train_steps: 100000 num_warmup_steps: 1000 optimizer: betas: - 0.9 - 0.95 eps: 1.0e-08 weight_decay: 1.0e-10 save_format: safetensors save_interval: 5000 scheduler_specific_kwargs: warmup_start_factor: 0.0 wandb_entity: null wandb_project: TurboVLA-RoboTwin