========================================================================
 HOW TO ACTUALLY RUN THE ROBOT  (pose6daug GR00T-N1.5, Franka EEF rig)
========================================================================

!! This MOVES the physical Franka arm. Have the e-stop in hand and the
!! workspace clear before you launch.

------------------------------------------------------------------------
0) WHICH COMMAND DOES WHAT  (read this first)
------------------------------------------------------------------------
  full_inference_mock.py   -> DOES NOT MOVE THE ROBOT. Robot + cameras are
                              mocked. Verification only. Safe to run anytime:

     PYTHONPATH=/home/jimin/myGR00T CUDA_VISIBLE_DEVICES=0 STEPS=48 \
       /home/jimin/miniconda3/envs/gr00t/bin/python \
       /home/jimin/jonghoon/mygr00t_franka_deploy/full_inference_mock.py

  deploy_franka_eef.py     -> MOVES THE REAL ROBOT. This is the one below.

------------------------------------------------------------------------
1) BEFORE YOU LAUNCH  (checklist)
------------------------------------------------------------------------
  [ ] Franka powered on, brakes released, e-stop within reach.
  [ ] Workspace clear; nothing in the arm's path.
  [ ] Both ZED cameras connected. Confirm the serials match:
        exterior = 37149196     wrist = 11845904
      (if different, edit --left_camera_id / --wrist_camera_id below)
  [ ] The DROID robot server/NUC is up (the `droid` RobotEnv connects to it).
  [ ] GPU 0 is free.

------------------------------------------------------------------------
2) RUN IT  (easiest — the example script)
------------------------------------------------------------------------
  bash /home/jimin/jonghoon/mygr00t_franka_deploy/run_example.sh

  - Task is preset to: "pick up the object and place it in the white basket"
  - Per-step translation clamp: 0.03  (safety)
  - Press Ctrl+C to stop the rollout early at any time.

------------------------------------------------------------------------
3) RUN IT  (explicit command — same thing, edit inline)
------------------------------------------------------------------------
  PYTHONPATH=/home/jimin/myGR00T CUDA_VISIBLE_DEVICES=0 \
  /home/jimin/miniconda3/envs/gr00t/bin/python -u \
    /home/jimin/jonghoon/mygr00t_franka_deploy/deploy_franka_eef.py \
    --task "pick up the object and place it in the white basket" \
    --max_trans_per_step 0.03 \
    --embodiment_tag=new_embodiment \
    --data_config=droid_franka_eef_real \
    --action_space=cartesian_position \
    --left_camera_id=37149196 \
    --wrist_camera_id=11845904 \
    --inference_head=main \
    --denoising_steps=4 \
    --max-timesteps=1500 \
    --logging_path=/home/jimin/droid/rollout_results/pose6daug-iclr120-20k \
    policy:checkpoint \
    --policy.config=franka_real_eef \
    --policy.dir=/home/jimin/pose6daug-gr00t-iclr120-20k/checkpoint-20000

  Omit --task to be prompted "Enter instruction:" interactively instead.
  Swap checkpoint-20000 for -5000/-10000/-15000 to try earlier steps.

------------------------------------------------------------------------
4) CONTROL MODE  --  CONFIRMED FROM THE CHECKPOINT
------------------------------------------------------------------------
  Checked the training normalization stats in the checkpoint metadata:
  the ACTION distribution == the STATE distribution (EEF x,y,z spanning the
  ~0.3-0.6 m workspace, not centered on 0). So this model outputs:

        ABSOLUTE cartesian target positions, in meters
        (NOT relative/delta, NOT velocity)

  Therefore:
     * use   --action_space=cartesian_position   (as in sections 2/3).
     * --max_trans_per_step 0.03 is a SAFETY MOVE LIMIT: it caps
       (target - current_EEF) to +/- 0.03 m per step, so the arm walks
       toward each predicted target no faster than 3 cm/step. It does NOT
       clamp the absolute target (that would drag the arm toward 0,0,0).
       (This was fixed in deploy_franka_eef.py; verified in the mock.)
     * do NOT use cartesian_velocity for this checkpoint.

  Tip: on the first real run, set a small --max_trans_per_step (e.g. 0.02),
  watch a few steps, then raise it.

------------------------------------------------------------------------
5) DO NOT
------------------------------------------------------------------------
  - Do NOT pass --resize_video : this config's VideoToTensor needs native
    720x1280 frames (it crops/resizes to 224 internally). Passing 192,320
    will crash it.
  - Do NOT run full_inference_mock.py expecting motion (see section 0).

========================================================================