| { |
| "policy": { |
| "policy_config": { |
| "_target_": "lerobot.common.policies.qwen3vl_policy.configuration_qwen3vl_policy.Qwen3VLPolicyConfig", |
| "max_state_dim": 50, |
| "max_action_dim": 50, |
| "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_projection": "linear", |
| "num_trainable_vlm_layers": null, |
| "proj_width": 1280, |
| "precision": "bfloat16", |
| "normalization_mode": "mean_std", |
| "model_mode": "flow_matching", |
| "image_shape": [ |
| 448, |
| 448 |
| ], |
| "image_keys": [ |
| "base_0_rgb", |
| "left_wrist_0_rgb", |
| "right_wrist_0_rgb" |
| ], |
| "add_control_mode": true, |
| "add_embodiment_name": true, |
| "num_steps": 10, |
| "n_action_steps": 50, |
| "mask_padded_actions": true, |
| "hold_at_subtask_boundary": false, |
| "hold_at_subtask_boundary_loss_fraction": 1.0, |
| "mask_temporally_padded_actions": true, |
| "state_noise_amplitude": 0.1, |
| "state_dropout_prob": 0.5, |
| "apply_noise_to_state_for_flow_matching": true, |
| "action_head_batch_multiplier": 1, |
| "add_state_proj_to_action_expert": true, |
| "enable_learnable_layer_combination": true, |
| "layer_combination_init": "identity", |
| "expert_block_stride": 4, |
| "suffix_attention_mask": "causal", |
| "attention_implementation": "flash_attention_2", |
| "freeze_vlm": false, |
| "freeze_vision_encoder": false, |
| "compile_sample_actions": false, |
| "is_knowledge_insulation": false, |
| "ce_loss_weight": 0.5, |
| "cognition_layer_mix": true, |
| "cognition_layer_mix_per_token": true, |
| "cognition_layer_mix_init": "uniform", |
| "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, |
| "tokenizer_max_length": 768, |
| "n_obs_steps": 1, |
| "obs_stride_seconds": 0.0, |
| "n_state_obs_steps": 1, |
| "state_obs_stride_seconds": 0.0, |
| "add_state_vlm": false, |
| "state_history_dropout_prob": 0.0, |
| "inputs_embeds_gradient_checkpointing": false |
| }, |
| "policy": { |
| "_target_": "lerobot.common.policies.qwen3vl_policy.modeling_qwen3vl_policy.Qwen3VLPolicy", |
| "config": { |
| "_target_": "lerobot.common.policies.qwen3vl_policy.configuration_qwen3vl_policy.Qwen3VLPolicyConfig", |
| "max_state_dim": 50, |
| "max_action_dim": 50, |
| "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_projection": "linear", |
| "num_trainable_vlm_layers": null, |
| "proj_width": 1280, |
| "precision": "bfloat16", |
| "normalization_mode": "mean_std", |
| "model_mode": "flow_matching", |
| "image_shape": [ |
| 448, |
| 448 |
| ], |
| "image_keys": [ |
| "base_0_rgb", |
| "left_wrist_0_rgb", |
| "right_wrist_0_rgb" |
| ], |
| "add_control_mode": true, |
| "add_embodiment_name": true, |
| "num_steps": 10, |
| "n_action_steps": 50, |
| "mask_padded_actions": true, |
| "hold_at_subtask_boundary": false, |
| "hold_at_subtask_boundary_loss_fraction": 1.0, |
| "mask_temporally_padded_actions": true, |
| "state_noise_amplitude": 0.1, |
| "state_dropout_prob": 0.5, |
| "apply_noise_to_state_for_flow_matching": true, |
| "action_head_batch_multiplier": 1, |
| "add_state_proj_to_action_expert": true, |
| "enable_learnable_layer_combination": true, |
| "layer_combination_init": "identity", |
| "expert_block_stride": 4, |
| "suffix_attention_mask": "causal", |
| "attention_implementation": "flash_attention_2", |
| "freeze_vlm": false, |
| "freeze_vision_encoder": false, |
| "compile_sample_actions": false, |
| "is_knowledge_insulation": false, |
| "ce_loss_weight": 0.5, |
| "cognition_layer_mix": true, |
| "cognition_layer_mix_per_token": true, |
| "cognition_layer_mix_init": "uniform", |
| "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, |
| "tokenizer_max_length": 768, |
| "n_obs_steps": 1, |
| "obs_stride_seconds": 0.0, |
| "n_state_obs_steps": 1, |
| "state_obs_stride_seconds": 0.0, |
| "add_state_vlm": false, |
| "state_history_dropout_prob": 0.0, |
| "inputs_embeds_gradient_checkpointing": false |
| } |
| } |
| }, |
| "image_augmentations": { |
| "base_camera": { |
| "_target_": "torchvision.transforms.v2.Compose", |
| "transforms": [ |
| { |
| "_target_": "torchvision.transforms.v2.Resize", |
| "size": [ |
| 448, |
| 448 |
| ], |
| "antialias": true |
| }, |
| { |
| "_target_": "torchvision.transforms.v2.RandomRotation", |
| "degrees": 1.0 |
| }, |
| { |
| "_target_": "torchvision.transforms.v2.RandomResizedCrop", |
| "size": [ |
| 448, |
| 448 |
| ], |
| "scale": [ |
| 0.97, |
| 1.0 |
| ], |
| "ratio": [ |
| 0.98, |
| 1.02 |
| ], |
| "antialias": true |
| }, |
| { |
| "_target_": "torchvision.transforms.v2.ColorJitter", |
| "brightness": 0.08, |
| "contrast": 0.08, |
| "saturation": 0.05, |
| "hue": 0.01 |
| } |
| ] |
| }, |
| "wrist_camera": { |
| "_target_": "torchvision.transforms.v2.Compose", |
| "transforms": [ |
| { |
| "_target_": "torchvision.transforms.v2.Resize", |
| "size": [ |
| 448, |
| 448 |
| ], |
| "antialias": true |
| }, |
| { |
| "_target_": "torchvision.transforms.v2.RandomRotation", |
| "degrees": 1.0 |
| }, |
| { |
| "_target_": "torchvision.transforms.v2.RandomResizedCrop", |
| "size": [ |
| 448, |
| 448 |
| ], |
| "scale": [ |
| 0.98, |
| 1.0 |
| ], |
| "ratio": [ |
| 0.99, |
| 1.01 |
| ], |
| "antialias": true |
| }, |
| { |
| "_target_": "torchvision.transforms.v2.ColorJitter", |
| "brightness": 0.06, |
| "contrast": 0.06, |
| "saturation": 0.04, |
| "hue": 0.01 |
| } |
| ] |
| } |
| }, |
| "robotics_dataset": { |
| "_target_": "lerobot.common.datasets.data_config.MixtureDataConfigFactory", |
| "name": "scripted_simxr_20260904_s0s1_step1_state_action_subtasks_shared", |
| "repo_id": "scripted_simxr_20260904", |
| "data_configs": [ |
| { |
| "_target_": "lerobot.common.datasets.data_config.LeRobotHumanoidS0S1DataConfig", |
| "name": "scripted_vla_s0s1_step1_state_action_subtasks_shared", |
| "asset_id": "scripted_simxr_20260904_s0s1_step1_state_action_shared", |
| "repo_id": "scripted_vla", |
| "root_dir": "/mnt/datasets/iefremov/scripted/vla", |
| "assets": { |
| "assets_dir": "/mnt/virtual_ai0001071-01239_SR006-nfs2/nerus/assets_rl", |
| "asset_id": "scripted_simxr_20260904_s0s1_step1_state_action_shared" |
| }, |
| "action_sequence_keys": [ |
| "action" |
| ], |
| "action_horizon": 50, |
| "action_offset": 0, |
| "action_sample_step": 1, |
| "control_mode": "joint", |
| "smooth_actions": true, |
| "has_velocities": true, |
| "use_stationary_masks": true, |
| "map_to_unified_space": false, |
| "map_to_closed_kinematic": true, |
| "input_kinematic_space": "open", |
| "model_kinematic_space": "closed", |
| "map_old_wrists_to_new": false, |
| "map_inspire_to_leadshipe": false, |
| "return_subtasks": true, |
| "return_subtasks_mode": "strict", |
| "prompt_from_subtask": true, |
| "exclude_negative_subtasks": false, |
| "base_config": { |
| "_target_": "lerobot.common.datasets.data_config.DataConfig", |
| "local_files_only": true, |
| "prompt_from_task": false |
| } |
| }, |
| { |
| "_target_": "lerobot.common.datasets.data_config.LeRobotHumanoidS0S1DataConfig", |
| "name": "simxr_merged_all_20260904_s0s1_step1_state_action_subtasks_shared", |
| "asset_id": "scripted_simxr_20260904_s0s1_step1_state_action_shared", |
| "repo_id": "simxr_merged_all_20260904", |
| "root_dir": "/mnt/virtual_ai0001071-01239_SR006-nfs2/iefremov/data/simxr_merged_all_20260904", |
| "validation_episodes": "/mnt/virtual_ai0001071-01239_SR006-nfs2/iefremov/data/simxr_merged_all_20260904/meta/validation_episodes.json", |
| "assets": { |
| "assets_dir": "/mnt/virtual_ai0001071-01239_SR006-nfs2/nerus/assets_rl", |
| "asset_id": "scripted_simxr_20260904_s0s1_step1_state_action_shared" |
| }, |
| "action_sequence_keys": [ |
| "action" |
| ], |
| "action_horizon": 50, |
| "action_offset": 0, |
| "action_sample_step": 1, |
| "control_mode": "joint", |
| "smooth_actions": true, |
| "has_velocities": true, |
| "use_stationary_masks": true, |
| "map_to_unified_space": false, |
| "map_to_closed_kinematic": true, |
| "input_kinematic_space": "open", |
| "model_kinematic_space": "closed", |
| "map_old_wrists_to_new": false, |
| "map_inspire_to_leadshipe": false, |
| "return_subtasks": true, |
| "return_subtasks_mode": "strict", |
| "prompt_from_subtask": true, |
| "exclude_negative_subtasks": false, |
| "base_config": { |
| "_target_": "lerobot.common.datasets.data_config.DataConfig", |
| "local_files_only": true, |
| "prompt_from_task": false |
| } |
| } |
| ], |
| "weights": [ |
| 10745464, |
| 4423926 |
| ] |
| }, |
| "optimizer": { |
| "_target_": "lerobot.common.optim.optimizers.AdamWConfig", |
| "lr": 5e-05, |
| "betas": [ |
| 0.9, |
| 0.95 |
| ], |
| "eps": 1e-08, |
| "weight_decay": 1e-08, |
| "grad_clip_norm": 1.0 |
| }, |
| "scheduler": { |
| "_target_": "lerobot.common.optim.schedulers.WarmupStableDecaySchedulerConfig", |
| "num_warmup_steps": 992, |
| "stable_phase_steps": 11851, |
| "num_decay_steps": 1971, |
| "peak_lr": 5e-05, |
| "decay_lr": 0.0 |
| }, |
| "training": { |
| "batch_size": 8, |
| "steps": 14814, |
| "num_warmup_steps": 992, |
| "num_decay_steps": null, |
| "log_freq": 10, |
| "save_freq": 14814, |
| "num_workers": 16, |
| "num_workers_validation": 8, |
| "shuffle": true, |
| "seed": 42, |
| "resume": false, |
| "prefetch_factor": 4, |
| "output_dir": "/mnt/virtual_ai0001071-01239_SR006-nfs2/nerus/training_runs/postnikov_qwen35_08b_scripted_simxr_20260904_step1_verylight_state_action_subtasks_sharedstats_nogc_1epoch_fastkernels_32gpu_2026-09-05_00-02", |
| "grad_accum_steps": 4, |
| "find_unused_parameters": true, |
| "use_fsdp": false, |
| "use_fsdp2": true, |
| "fsdp_sharding": "full", |
| "gradient_checkpointing_enabled": false, |
| "activation_offload_enabled": false, |
| "activation_gpu_limit_gb": 0.0, |
| "compile": false, |
| "enable_validation": true, |
| "use_validation_list": true, |
| "val_split": 0.03, |
| "val_freq": 14814, |
| "max_val_steps": 100, |
| "recompute_norm_stats": false, |
| "visualization": { |
| "plots_per_validation": 5, |
| "subplot_height": 3, |
| "subplot_width": 8 |
| }, |
| "fsdp_keep_trainable_vlm_unsharded": false, |
| "validate_at_start": false, |
| "validation_batch_size": 1, |
| "kernel_prewarm": true, |
| "save_checkpoint": true |
| }, |
| "logging": { |
| "clearml": { |
| "task": { |
| "project_name": "Manipulation/2D Manipulation/lerobot", |
| "task_name": "postnikov_qwen35_08b_scripted_simxr_20260904_step1_verylight_state_action_subtasks_sharedstats_nogc_1epoch_fastkernels_32gpu_2026-09-05_00-02", |
| "task_type": "training" |
| }, |
| "logging": { |
| "level": "INFO" |
| }, |
| "upload_method": "rclone", |
| "s3_output_uri": "s3://obs.ru-moscow-1.hc.sbercloud.ru:443/amr-usr-manip/checkpoints", |
| "keep_latest_local": 1 |
| } |
| }, |
| "research_ledger_mode": "ineligible", |
| "run_name": "postnikov_qwen35_08b_scripted_simxr_20260904_step1_verylight_state_action_subtasks_sharedstats_nogc_1epoch_fastkernels_32gpu_2026-09-05_00-02", |
| "tags": [ |
| "postnikov-aij", |
| "qwen3.5", |
| "qwen35-0.8b", |
| "scripted-simxr-mix", |
| "shared-norm-stats", |
| "state-action-target", |
| "next-state-target", |
| "subtask-prompt", |
| "no-gradient-checkpointing", |
| "s0s1", |
| "step1", |
| "very-light-aug", |
| "no-history", |
| "from-scratch", |
| "one-epoch", |
| "fast-kernels" |
| ], |
| "assets_dir": "/mnt/virtual_ai0001071-01239_SR006-nfs2/nerus/assets_rl", |
| "pretrained_path": null, |
| "model_id": null |
| } |
