Action space mismatch across paper / dataset / code (20 vs 21 vs 24 vs 25)

[mxlin03]

Hi, I’m confused by inconsistent action-space definitions:

• Paper body (“Unified observation and action space”): 16 binary buttons + 4 joystick axes = 20 dims
• Paper appendix (A.1): action chunk a ∈ R^(16×24) (24 dims per step)
• Dataset parquet: 17 boolean button columns (includes guide) + (j_left, j_right) each with (x,y) = 21 dims
• Code/checkpoint: model uses 25 dims = 21 button tokens + 4 joystick axes, where the extra 4 tokens are:
  RIGHT_UP, RIGHT_BOTTOM, RIGHT_LEFT, RIGHT_RIGHT

Questions:

1. What is the canonical per-step action dim used by NitroGen (20/21/24/25)?
2. How is the dataset 21-dim action mapped to the model 25-dim action (esp. discretization of right stick into RIGHT_* tokens: thresholds/dead-zone)?

[loicmagne]

Hi @mxlin03, thanks for pointing that out

As you noted the full action space used by our model has 25 dimensions with 4 (2+2) joystick axes and 21 buttons:

BUTTON_ACTION_TOKENS = [
    'DPAD_DOWN', 'DPAD_LEFT', 'DPAD_RIGHT', 'DPAD_UP',
    'LEFT_SHOULDER', 'LEFT_THUMB', 'LEFT_TRIGGER',
    'RIGHT_SHOULDER', 'RIGHT_THUMB', 'RIGHT_TRIGGER',
    'SOUTH', 'WEST', 'EAST', 'NORTH',
    'BACK',  'START',
    'RIGHT_BOTTOM', 'RIGHT_LEFT', 'RIGHT_RIGHT', 'RIGHT_UP',
    'GUIDE',
]

RIGHT_BOTTOM RIGHT_LEFT RIGHT_RIGHT RIGHT_UP and GUIDE ended up not being used in neither the model or dataset, they're artifacts. This is why we (inconsistently) removed them from the action dimension count in the paper, we'll try to make it clearer. You can just set these values to 0 if you want to fine-tune the model