chore: canonical naming migration

rskill.yaml

@@ -13,12 +13,12 @@
 
 # ── Identity ───────────────────────────────────────────────────────────────
 schema_version: "0.1"
-name: "OpenRAL/rskill-smolvla-libero"
-evaluated_tasks: ["libero_spatial", "libero_object", "libero_goal", "libero_10"]  # ADR-0060: trained on the full LIBERO set (all 4 suites, 40 tasks)
+name: "OpenRAL/rskill-smolvla-franka_panda-libero_spatial-bf16"
+evaluated_tasks: ["libero_spatial", "libero_object", "libero_goal", "libero_10"]  # trained on the full LIBERO set (all 4 suites, 40 tasks)
 version: "0.1.0"
 license: "apache-2.0"
 role: "s1"
-kind: "vla"  # ADR-00XX: rSkill kind discriminator. "vla" = learnable Vision-Language-Action policy.
+kind: "vla"  # rSkill kind discriminator. "vla" = learnable Vision-Language-Action policy.
 
 # ── Policy identity ────────────────────────────────────────────────────────
 model_family: "smolvla"
@@ -27,7 +27,7 @@ model_family: "smolvla"
 # LIBERO benchmark targets table-top manipulation with a Franka Panda
 # arm — the canonical embodiment used under
 # `openral sim run --config smolvla_libero_spatial.yaml` (matches the
-# pi05-libero-nf4 / xvla-libero manifests).
+# pi05-libero-int8 / xvla-libero manifests).
 embodiment_tags:
   - "franka_panda"
 
@@ -44,7 +44,7 @@ sensors_required:
     min_width: 224
     min_height: 224
 
-# Output side (ADR-0013). For the canonical franka_panda embodiment the
+# Output side. For the canonical franka_panda embodiment the
 # loader auto-fills n_dof + vla_action_key from robots/franka_panda/robot.yaml.
 actuators_required:
   - kind: "joint_position"
@@ -56,15 +56,16 @@ runtime: "pytorch"
 quantization:
   dtype: "bf16"
   backend: "pytorch"
-# ADR-0077 — VRAM footprint per dtype, so the deploy can verify this VLA + its
+# VRAM footprint per dtype, so the deploy can verify this VLA + its
 # paired reward model co-reside on the GPU before loading. bf16 measured at 0.93 GB
 # weights at load (RTX 4070); 1.2 GB declared with inference-activation headroom.
 min_vram_gb:
   bf16: 1.2
-# ADR-0077 — a VLA emits no success signal of its own, so it runs with a reward /
-# progress monitor resident alongside it (ADR-0074). Robometer (NF4, 3.6 GB) is the
-# default; 1.2 + 3.6 = 4.8 GB fits an 8 GB card with room to spare.
-reward_rskill_name: "OpenRAL/rskill-robometer-4b-nf4"
+# A VLA emits no success signal of its own, so it runs with a reward /
+# progress monitor resident alongside it as the model's completion signal.
+# Robometer (NF4, 3.6 GB) is the default; 1.2 + 3.6 = 4.8 GB fits an 8 GB
+# card with room to spare.
+reward_rskill_name: "OpenRAL/rskill-robometer_4b-any-general-nf4"
 # Pin to the HEAD SHA at packaging time for reproducibility (principle 8).
 # Update by re-running: tools/rskill_publisher.py --bump-revision
 weights_uri: "hf://lerobot/smolvla_libero"
@@ -85,13 +86,13 @@ image_preprocessing:
     top: "image"
     wrist: "image2"
 state_contract:
-  # ADR-0027 — the LIBERO checkpoint trains on 8-D TASK-SPACE proprio
+  # the LIBERO checkpoint trains on 8-D TASK-SPACE proprio
   # (eef_pos ‖ eef_axisangle ‖ gripper). The benchmark (`openral sim run`)
   # supplies it directly; deploy (`openral deploy sim`) assembles it from live
   # TF + JointState via the `libero_eef8d` state-adapter. Without the layout the
   # runner falls back to raw joint-space state and the policy never reaches the
   # target. The franka deploy sim publishes a static `world -> panda_link0` from
-  # the sim's robot mount pose (sim_sensor_bridge, ADR-0027) so the EE pose is
+  # the sim's robot mount pose (sim_sensor_bridge) so the EE pose is
   # read in the same world frame as robosuite's `robot0_eef_pos`.
   layout: "libero_eef8d"
   dim: 8
@@ -142,7 +143,7 @@ description: >
   here matches the paper's reported configuration on all five
   architecture fields — see header comment for the rejected sibling.
 
-# ADR-0022 — action vocabulary surfaced to the reasoner LLM tool
+# Action vocabulary surfaced to the reasoner LLM tool
 # palette so it can pick this skill by what it does (action verb +
 # object + scene), not just by its slug.
 actions:
@@ -159,11 +160,11 @@ scenes:
   - "tabletop"
   - "kitchen"
 
-# ADR-0019 — per-checkpoint action contract (consumed by the dataset bridge
+# Per-checkpoint action contract (consumed by the dataset bridge
 # to bind the LeRobot v3 `action` feature shape).
 action_contract:
   dim: 7
-  # ADR-0036 — LIBERO action space is 6-D OSC end-effector delta + gripper.
+  # LIBERO action space is 6-D OSC end-effector delta + gripper.
   # Declaring the representation makes deploy-sim dispatch cartesian_delta +
   # gripper (via canonical_slots_for_representation) instead of defaulting the
   # whole vector to JOINT_POSITION, which the joint-space envelope rejects.