config.json

{
  "type": "vla_jepa",
  "n_obs_steps": 1,
  "input_features": {
    "observation.images.image": {
      "type": "VISUAL",
      "shape": [
        3,
        224,
        224
      ]
    }
  },
  "output_features": {
    "action": {
      "type": "ACTION",
      "shape": [
        7
      ]
    }
  },
  "device": null,
  "use_amp": false,
  "use_peft": false,
  "push_to_hub": true,
  "repo_id": null,
  "private": null,
  "tags": null,
  "license": null,
  "pretrained_path": null,
  "chunk_size": 7,
  "n_action_steps": 7,
  "normalization_mapping": {
    "VISUAL": "IDENTITY",
    "STATE": "MEAN_STD",
    "ACTION": "MIN_MAX"
  },
  "qwen_model_name": "Qwen/Qwen3-VL-2B-Instruct",
  "jepa_encoder_name": "facebook/vjepa2-vitl-fpc64-256",
  "freeze_qwen": false,
  "enable_world_model": true,
  "reinit_modules": null,
  "tokenizer_padding_side": "left",
  "prompt_template": "Your task is {instruction}. Infer the temporal dynamics from frames {actions} and produce the corresponding policy actions {e_actions}.",
  "special_action_token": "<|action_{}|>",
  "embodied_action_token": "<|embodied_action|>",
  "action_dim": 7,
  "state_dim": 8,
  "num_action_tokens_per_timestep": 8,
  "num_embodied_action_tokens_per_instruction": 32,
  "num_inference_timesteps": 4,
  "action_hidden_size": 1024,
  "action_model_type": "DiT-B",
  "action_num_layers": 16,
  "action_num_heads": 12,
  "action_attention_head_dim": 64,
  "action_dropout": 0.2,
  "action_num_timestep_buckets": 1000,
  "action_noise_beta_alpha": 1.5,
  "action_noise_beta_beta": 1.0,
  "action_noise_s": 0.999,
  "num_target_vision_tokens": 32,
  "action_max_seq_len": 1024,
  "num_video_frames": 8,
  "predictor_depth": 12,
  "predictor_num_heads": 8,
  "predictor_mlp_ratio": 4.0,
  "predictor_dropout": 0.0,
  "world_model_loss_weight": 0.1,
  "jepa_tubelet_size": 2,
  "repeated_diffusion_steps": 8,
  "resize_images_to": [
    224,
    224
  ],
  "binarize_gripper_action": true,
  "pre_snap_gripper_action": true,
  "clip_normalized_actions": true,
  "torch_dtype": "bfloat16",
  "optimizer_lr": 0.0001,
  "optimizer_betas": [
    0.9,
    0.95
  ],
  "optimizer_eps": 1e-08,
  "optimizer_weight_decay": 1e-08,
  "optimizer_grad_clip_norm": 1.0,
  "scheduler_warmup_steps": 5000,
  "scheduler_decay_steps": 30000,
  "scheduler_decay_lr": 1e-05,
  "gripper_dim": 6,
  "gripper_threshold": 0.5
}