rskill.yaml: sync to OpenRAL namespace + current schema (joint_units, evaluated_tasks, ADR-0071)

rskill.yaml

@@ -33,11 +33,18 @@
 # `task.max_steps` to MS3's `max_episode_steps` so the rollout isn't
 # silently truncated at MS3's default 50 steps.
 #
-# The upstream model repo doesn't ship `policy_*processor.json`. The
-# SmolVLA adapter (also this PR) detects that and rebuilds the
-# lerobot processors from `dataset_uri`'s `meta/episodes_stats.jsonl`
-# — same path applies to any community finetune uploaded without
-# processors.
+# The upstream model repo (Calvert0921/smolvla_franka_liftcube_1000)
+# ships only config.json + model.safetensors + train_config.json — no
+# `policy_*processor.json`. Rather than rely on the SmolVLA adapter's
+# dataset-stats fallback at every load (a 404 + recompute against
+# `dataset_uri`'s `meta/episodes_stats.jsonl`), the processor pair was
+# rebuilt once via `make_pre_post_processors(policy.config,
+# dataset_stats=...)` and uploaded to THIS rSkill's own HF repo
+# (OpenRAL/rskill-smolvla-maniskill-franka). The rebuilt normalizer
+# stats match the checkpoint's baked `normalize_inputs` /
+# `normalize_targets` buffers exactly, so the processors are faithful to
+# how the model was trained. The fallback path still exists for any
+# other community finetune uploaded without processors.
 
 # ── Identity ───────────────────────────────────────────────────────────────
 schema_version: "0.1"
@@ -87,14 +94,15 @@ weights_uri: "hf://Calvert0921/smolvla_franka_liftcube_1000"
 
 # ── Preprocessing (all knobs needed to interpret IO) ───────────────────────
 # Per-file URIs for the lerobot PolicyProcessorPipeline. The upstream
-# repo only ships the safetensors + config.json + train_config.json (no
-# stored `policy_*processor.json`); the SmolVLA adapter falls back to
-# building processors from `policy.config` via `make_pre_post_processors`
-# at load time. These URIs are left pointing at the canonical filenames
-# so a future re-upload that includes them is picked up automatically.
+# weights repo ships no `policy_*processor.json`, so the pair (plus the
+# normalizer/unnormalizer `*.safetensors` state files referenced by the
+# JSON `steps`) was rebuilt from the checkpoint config + training-dataset
+# stats and uploaded to this rSkill's own HF repo. `materialize_processor_dir`
+# downloads exactly these files (and their `state_file` siblings) — no 404,
+# no dataset-stats recompute at load time.
 processors:
-  preprocessor_uri: "hf://Calvert0921/smolvla_franka_liftcube_1000/policy_preprocessor.json"
-  postprocessor_uri: "hf://Calvert0921/smolvla_franka_liftcube_1000/policy_postprocessor.json"
+  preprocessor_uri: "hf://OpenRAL/rskill-smolvla-maniskill-franka/policy_preprocessor.json"
+  postprocessor_uri: "hf://OpenRAL/rskill-smolvla-maniskill-franka/policy_postprocessor.json"
 # Per-checkpoint knobs the processor JSONs don't encode. The model was
 # trained on raw SAPIEN renders (no 180° flip); cameras need an alias
 # remap from the in-tree scene-side keys (camera1 / camera2) to the
@@ -102,8 +110,8 @@ processors:
 image_preprocessing:
   flip_180: false
   aliases:
-    camera1: "up"
-    camera2: "wrist"
+    top: "up"
+    wrist: "wrist"
 state_contract:
   dim: 9
 
@@ -122,6 +130,19 @@ latency_budget:
 paper_url: "https://arxiv.org/abs/2506.01844"
 dataset_uri: "hf://Calvert0921/SmolVLA_LiftCube_Franka_1000"
 source_repo: "hf://Calvert0921/smolvla_franka_liftcube_1000"
+# Task-data gate (ADR-0060): this checkpoint was trained on the *LiftCube*
+# dataset (a ManiSkill2-style "lift the cube to a height" task) which has NO
+# equivalent env in ManiSkill3 3.0.1 — MS3 ships PickCube-v1 (grasp + place at
+# a goal + stay static), a different task. Declaring the (MS3-absent) true task
+# makes the benchmark runner refuse every real MS3 scene with
+# ROSCapabilityMismatch. Verified empirically 2026-06-19: retargeted to
+# PickCube-v1 it runs (50 steps, ~15 ms/step real inference) and attempts the
+# cube but cannot satisfy PickCube's goal-placement success (success=False) —
+# exactly the plausible-but-unsolvable rollout the gate exists to block. This
+# rSkill therefore has no valid MS3 benchmark; replace with a PickCube-trained
+# policy (e.g. the author's PegInsertion checkpoint -> PegInsertionSide-v1, or
+# an MS3 OpenVLA-OFT policy) before pairing it with a maniskill scene.
+evaluated_tasks: ["maniskill3/LiftCube"]
 
 description: >
   SmolVLA (0.45 B, lerobot/smolvla_base) finetuned on
@@ -149,4 +170,9 @@ action_contract:
   dim: 8
 
 benchmarks:
-  maniskill3_franka_pick_cube: 0.1
+  # PickCube-v1 via the curated franka_panda suite `maniskill3_panda`
+  # (auto-filters to this rSkill's task). 0.0 = the shipped LiftCube
+  # checkpoint cannot satisfy PickCube's goal-placement success (task
+  # mismatch — see the MS3 LiftCube-gap note above); replace with a
+  # PickCube-trained policy to lift this.
+  maniskill3_panda: 0.0