{
  "format": "vpp2_release_config_v1",
  "model_type": "vpp2",
  "model_name": "Video Prediction Policy 2: Predict Better, Act Better",
  "base_model": "Wan-AI/Wan2.1-I2V-14B-480P",
  "code_repository": "https://github.com/roboterax/video-prediction-policy-2",
  "shared_assets": {
    "wan_model_dir": "checkpoints/Wan2.1-I2V-14B-480P"
  },
  "variants": {
    "video_stage1_49f": {
      "video_checkpoint": "checkpoints_video/vpp2-video-stage1-49f.pth",
      "num_frames": 49,
      "inference": {
        "sampling_steps": 30,
        "guidance_scale": 4,
        "sigma_shift": 3,
        "height": 240,
        "width": 416
      }
    },
    "video_stage2_17f": {
      "video_checkpoint": "checkpoints_video/vpp2-video-stage2-17f.pth",
      "num_frames": 17,
      "inference": {
        "sampling_steps": 30,
        "guidance_scale": 4,
        "sigma_shift": 3,
        "height": 240,
        "width": 416
      }
    },
    "robodojo_his10k": {
      "video_checkpoint": "checkpoints/initialization/robodojo_his10k.pt",
      "use": "History-conditioned Video-10k initializer for joint Video + Action2B training"
    },
    "robodojo_joint2b_100k": {
      "video_checkpoint": "checkpoints/joint2b_s100000/video.pt",
      "action_checkpoint": "checkpoints/joint2b_s100000/action.pt",
      "manifest": "checkpoints/joint2b_s100000/manifest.json",
      "dataset_stats": "checkpoints/joint2b_s100000/dataset_stats.json",
      "inference": {
        "steps": 10,
        "shift": 1,
        "horizon": 32,
        "replan": 24,
        "history_stride": 25,
        "seed": 1
      }
    },
    "libero_video10k_action30k": {
      "video_checkpoint": "checkpoints/libero/video_step010000.pt",
      "action_checkpoint": "checkpoints/libero/action_step030000.pt",
      "manifest": "checkpoints/libero/manifest.json",
      "dataset_stats": "checkpoints/libero/dataset_stats.json",
      "benchmarks": [
        "LIBERO",
        "LIBERO-OOD",
        "LIBERO-PRO"
      ],
      "inference": {
        "action_shift": 5,
        "euler_steps": 10,
        "action_horizon": 32,
        "execute_actions": 10,
        "tokenizer_max_len": 128,
        "normalization": "min/max"
      }
    }
  }
}