VLA/workspace/example/maniparena_example.yml

# ManipArena / CVPR dual-arm finetune example (6D relative action, 448px, 3 cameras).
#
# Replace every /path/to/* placeholder before training:
#
#   model.config_path           -> model architecture JSON (mot_flash_mask_causal_xloss.json)
#   model.processor_path        -> Qwen2.5-VL-3B-Instruct directory
#   model.pretrained_path       -> same as processor_path, or HuggingFace cache path
#   data.lerobot_config.repo_id -> local LeRobot dataset root (cvpr_4tasks)
#   data.norm_stats_path        -> q01/q99 normalization JSON
#   checkpoint.save_path        -> writable directory for training checkpoints
#   checkpoint.resume_from      -> Wall-OSS-0.5 pretrained .safetensors or checkpoint directory
#
# Compute norm stats first:
#
#   python scripts/compute_norm_stats.py \
#     --train_config workspace/example/maniparena_example.yml \
#     --data_root /path/to/cvpr_4tasks \
#     --output_path /path/to/cvpr_4tasks_norm_stats.json
#
# Launch training (from repo root):
#
#   torchrun --nproc_per_node=<N> wall_x/trainer/fsdp_trainer/train_fsdp.py \
#     --config workspace/example/maniparena_example.yml
#
# Strategy: keep dof / agent_pos totals at 26 to match the pretraining action space
# via ``action_padding``. The lerobot collator right-pads real action / state dims;
# loss does not flow through the padded tail.

model_type: qwen2_5

task:
  # Dual-arm 6D relative action: 10 + 10 = 20, plus action_padding(6) = 26.
  dof_config:
    follow_left_ee_cartesian_pos_relative: 3
    follow_left_ee_rotation_6D_relative: 6
    follow_left_gripper: 1
    follow_right_ee_cartesian_pos_relative: 3
    follow_right_ee_rotation_6D_relative: 6
    follow_right_gripper: 1
    action_padding: 6
  ar_dof_config:
    follow_left_ee_cartesian_pos_relative: 3
    follow_left_ee_rotation_6D_relative: 6
    follow_left_gripper: 1
    follow_right_ee_cartesian_pos_relative: 3
    follow_right_ee_rotation_6D_relative: 6
    follow_right_gripper: 1
    action_padding: 6
  agent_pos_config:
    follow_left_ee_cartesian_pos: 3
    follow_left_ee_rotation_6D: 6
    follow_left_gripper: 1
    follow_right_ee_cartesian_pos: 3
    follow_right_ee_rotation_6D: 6
    follow_right_gripper: 1
    action_padding: 6
  action_horizon: 32
  action_horizon_flow: 32
  use_state_string_representation: false

model:
  backbone: qwen2_5
  config_path: /path/to/wall-oss-0.5/config.json
  processor_path: /path/to/Qwen2.5-VL-3B-Instruct
  pretrained_path: /path/to/Qwen2.5-VL-3B-Instruct
  attn_deterministic: true
  use_ema: false
  flow_loss_weight: 1.0
  ar_loss_weight: 0.01

hyperparams:
  num_epoch: 100
  batch_size_per_gpu: 4
  gradient_accumulation_steps: 4
  seed: 10222
  optimizer:
    optimizer_type: adamw
    learning_rate: 5.0e-05
    max_grad_norm: 1.0
    enable_grad_clip: true
    betas: [0.9, 0.95]
    weight_decay: 1.0e-8
    eps: 1.0e-8
  scheduler:
    scheduler_type: cosine
    num_warmup_steps: 1000
    num_training_steps: 200000
    min_lr: 1.0e-6

distributed:
  use_fsdp: true
  use_mixed_precision: true
  bf16: true

data:
  dataset_type: lerobot
  lerobot_config:
    repo_id: /path/to/cvpr_4tasks
    root: null
  key_mappings:
    camera:
      observation.images.faceImg: face_view
      observation.images.leftImg: left_wrist_view
      observation.images.rightImg: right_wrist_view
    state: observation.state
    action: action
  norm_stats_path: /path/to/cvpr_4tasks_norm_stats.json
  train_test_split: 0.95
  num_workers: 4
  max_length: 1024
  resolution:
    face_view: 448
    left_wrist_view: 448
    right_wrist_view: 448

logging:
  log_name: maniparena_ft
  log_project: lerobot_maniparena_ft
  log_entity: your_wandb_entity
  use_wandb: true
  log_interval: 10
  save_interval: 2000
  val_interval: 1000000
  epoch_save_interval: 1

checkpoint:
  save_path: /path/to/cvpr_4tasks
  # Single-file .safetensors loads as pretrain weights before FSDP wrapping.
  # Use a checkpoint directory for full resume (optimizer / scheduler / RNG).
  resume_from: /path/to/wall-oss-0.5/model.safetensors

debug:
  profile: false
  nvtx: false