| { |
| "n_obs_steps": 1, |
| "device": "cuda", |
| "use_amp": false, |
| "normalization_mode": "mean_std", |
| "obs_stride_seconds": 0.0, |
| "n_state_obs_steps": 1, |
| "state_obs_stride_seconds": 0.0, |
| "n_action_steps": 50, |
| "model_type": "qwen3vlpolicy", |
| "num_steps": 10, |
| "max_state_dim": 50, |
| "max_action_dim": 50, |
| "map_to_unified_space": false, |
| "unified_space_dim": 64, |
| "add_action_space_factorization": true, |
| "image_keys": [ |
| "base_0_rgb", |
| "left_wrist_0_rgb", |
| "right_wrist_0_rgb" |
| ], |
| "tokenizer_max_length": 768, |
| "proj_width": 1280, |
| "expert_block_stride": 4, |
| "base_vlm_model": "Qwen/Qwen3.5-0.8B", |
| "cognition_mode": "frozen_vlm_context_queries", |
| "action_expert_model": "Qwen/Qwen3.5-0.8B", |
| "num_cognition_tokens": 64, |
| "cognition_vlm_layer": -1, |
| "cognition_layer_mix": true, |
| "cognition_layer_mix_per_token": true, |
| "cognition_layer_mix_init": "uniform", |
| "cognition_projection": "linear", |
| "cognition_context_width": 512, |
| "cognition_context_num_heads": 8, |
| "cognition_context_ffn_multiplier": 4, |
| "cognition_frozen_vlm_reference_kernels": false, |
| "cognition_action_expert_reference_kernels": false, |
| "num_trainable_vlm_layers": null, |
| "inputs_embeds_gradient_checkpointing": false, |
| "precision": "bfloat16", |
| "suffix_attention_mask": "causal", |
| "model_mode": "flow_matching", |
| "inference_mode": "flow_matching", |
| "image_shape": [ |
| 448, |
| 448 |
| ], |
| "default_temperature": 0.7, |
| "discrete_state_input": false, |
| "continuous_state_input": false, |
| "add_state_vlm": false, |
| "state_history_dropout_prob": 0.0, |
| "add_state_proj_to_action_expert": true, |
| "apply_noise_to_state_for_flow_matching": true, |
| "state_noise_amplitude": 0.1, |
| "state_dropout_prob": 0.5, |
| "attention_implementation": "flash_attention_2", |
| "mask_padded_actions": true, |
| "hold_at_subtask_boundary": false, |
| "hold_at_subtask_boundary_loss_fraction": 1.0, |
| "mask_temporally_padded_actions": true, |
| "enable_learnable_layer_combination": true, |
| "layer_combination_init": "identity", |
| "is_knowledge_insulation": false, |
| "ce_loss_weight": 0.5, |
| "freeze_vlm": false, |
| "compile_sample_actions": false, |
| "state_special_token_id": 228, |
| "clip_state": false, |
| "freeze_vision_encoder": false, |
| "use_cache": false, |
| "add_control_mode": true, |
| "add_embodiment_name": true, |
| "action_head_batch_multiplier": 1, |
| "type": "qwen3vlpolicy" |
| } |
