[Vertax42]

lerobot-xense renamed both the tactile knob and the tactile camera keys. This repo still spoke the old vocabulary, and two of the three call sites were already broken at runtime — this is a bug fix as much as a rename.

1. enable_tactile_sensors → enable_tactile

Upstream moved the knob onto the typed gripper: block at c49d4aa5 (what a gripper carries is a property of the gripper, not of the arm holding it). BiFlexivRizon4RTConfig no longer declares it, so every Flexiv run died in the recipe loader — reproduced before touching anything:

TypeError: BiFlexivRizon4RTConfig.__init__() got an unexpected keyword argument 'enable_tactile_sensors'

flexiv_recipe.load_robot_config() now routes each override to whichever dataclass declares it. Callers still pass a flat kwarg; the loader decides the nesting level, rebuilds the gripper block, and lets __post_init__ re-clone it per side. A name neither level knows is now an error, not a silent no-op:

forward-04.yaml: override(s) ['enable_tactile_sensors'] name no field on
BiFlexivRizon4RTConfig or its TaccapFollowerConfig block.

The override warning gained the level it landed on: gripper.enable_tactile: True -> False.

The ARX5 path is renamed to match — its lerobot config is renamed in the sister-repo PR (Vertax42/lerobot-xense, branch refactor/arx5-enable-tactile). Land that one first, or bi_arx5 inference breaks on the new kwarg.

2. <arm>_tactile_{0,1} → <arm>_tactile_<finger>

The suffix now names the jaw the pad sits on (serial parity, odd → left) rather than USB enumeration order, which was a stable order but not an identity.

bi_arx5's camera mapping still asked for left_tactile_0, a key lerobot stopped emitting at 1146d034. get_images() only warns on a missing camera — so tactile inference on ARX5 had been feeding the model nothing, one log line per step. There was a second latent crash behind it: real_env emitted left_tactile_0 while AlohaTactileInputs.EXPECTED_CAMERAS wanted cam_left_tactile, which fails the subset check. Both are gone: lerobot's key is carried through unchanged, robot → transform → model slot, with no renaming step anywhere.

This is a relabel, not a retrain

embed_prefix() selects a branch by name and applies one shared encoder per branch (PaliGemma.img / tactile_img), so no weight is keyed by image name. The only thing a rename can disturb is token order — which is sorted-key order, since JAX canonicalises dict keys. Verified it is unchanged:

base_0_rgb          -> base_0_rgb             same slot
left_wrist_0_rgb    -> left_wrist_0_rgb       same slot
right_wrist_0_rgb   -> right_wrist_0_rgb      same slot
left_tactile_0_rgb  -> left_tactile_left_rgb  same slot
right_tactile_0_rgb -> right_tactile_left_rgb same slot

preprocess_observation_tactile's "tactile" in key augmentation branch still matches.

The model consumes one pad per arm; this bench wires the left jaw of each (OG000337, OG000339 — both odd). The other two pads are recorded and unused. Feeding all four is a change of input shape, not of naming, and is not attempted here.

Interaction with #15

None, but worth stating: #15 selects tactile streams by substring (TACTILE_KEY_MARKER = "tactile"), which the new keys still match, and the two branches touch disjoint files. #15's docs and its _XENSE_VIDEO_KEYS test fixture still spell _tactile_{0,1}; those are cosmetic and can be swept in a follow-up.

Verification

• ruff check src/ examples/ clean
• pytest examples/ src/openpi/policies → 67 passed; the one failure (policy_rtc_test) reproduces on a clean tree
• All 5 Flexiv recipes decode and route enable_tactile=False to gripper / left_gripper / right_gripper
• BiARX5Config(enable_tactile=True) emits exactly the keys real_env now asks for

🤖 Generated with Claude Code