chore: publish rSkill OpenRAL/rskill-smolvla-maniskill-franka v0.1.0

README.md

@@ -18,7 +18,7 @@ language:
 
 > **OpenRAL rSkill** — SmolVLA (0.45 B) finetuned on a 1000-demo Franka
 > LiftCube dataset in ManiSkill3 SAPIEN, packaged for use with the
-> [OpenRAL](https://github.com/AdrianLlopart/openral) robot agent framework.
+> [OpenRAL](https://github.com/OpenRAL/openral) robot agent framework.
 
 This package wraps
 [`Calvert0921/smolvla_franka_liftcube_1000`](https://huggingface.co/Calvert0921/smolvla_franka_liftcube_1000)
@@ -73,7 +73,7 @@ this layout.
 
 | Robot | Embodiment tag | Status | Notes |
 | --- | --- | --- | --- |
-| Franka Panda (ManiSkill3 SAPIEN) | `franka_panda` | ✓ end-to-end | Manifest validates and `ral sim run --view` produces a live SAPIEN window of the policy lifting the cube; processors are auto-synthesized from the training dataset's `meta/episodes_stats.jsonl` because the upstream model repo doesn't ship `policy_*processor.json`. |
+| Franka Panda (ManiSkill3 SAPIEN) | `franka_panda` | ✓ end-to-end | Manifest validates and `openral sim run --view` produces a live SAPIEN window of the policy lifting the cube; processors are auto-synthesized from the training dataset's `meta/episodes_stats.jsonl` because the upstream model repo doesn't ship `policy_*processor.json`. |
 
 ## Sensors required
 
@@ -91,7 +91,7 @@ Full schema: [`openral_core.schemas.RSkillManifest`](../../python/core/src/openr
 
 | Field | Value |
 | --- | --- |
-| `name` | `AdrianLlopart/rskill-smolvla-maniskill-franka` |
+| `name` | `OpenRAL/rskill-smolvla-maniskill-franka` |
 | `version` | `0.1.0` |
 | `license` | `apache-2.0` |
 | `role` | `s1` (fast visuomotor policy) |
@@ -122,7 +122,7 @@ just bootstrap && uv sync --all-packages --group sim --group maniskill3
 
 # End-to-end rollout (live SAPIEN window via --view).
 DISPLAY=:1 uv run --group sim --group maniskill3 \
-    ral sim run --config examples/sim/smolvla_maniskill_pick_cube.yaml \
+    openral sim run --config scenes/benchmarks/smolvla_maniskill_pick_cube.yaml \
                 --rskill rskill://rskills/smolvla-maniskill-franka \
                 --view
 ```
@@ -135,11 +135,11 @@ viewer. On a warm cache the policy lifts the cube within ~200 steps
 ## Evaluation
 
 No benchmarks shipped yet (`eval/.gitkeep` only). The headline LiftCube
-success rate will be populated by `ral benchmark run` once a paired
+success rate will be populated by `openral benchmark run` once a paired
 benchmark suite lands in `benchmarks/`:
 
 ```bash
-ral benchmark run \
+openral benchmark run \
     --suite maniskill3_pick_place \
     --vla smolvla:rskill://rskills/smolvla-maniskill-franka
 ```
@@ -185,6 +185,6 @@ should be updated to match.
 ## See also
 
 - [`robots/franka_panda/robot.yaml`](../../robots/franka_panda/robot.yaml) — RobotDescription manifest.
-- [`examples/sim/smolvla_maniskill_pick_cube.yaml`](../../examples/sim/smolvla_maniskill_pick_cube.yaml) — paired SimEnvironment config.
-- [`examples/sim/maniskill3_pick_cube.yaml`](../../examples/sim/maniskill3_pick_cube.yaml) — bare-backend MS3 PickCube wiring test.
+- [`scenes/benchmarks/smolvla_maniskill_pick_cube.yaml`](../../scenes/benchmarks/smolvla_maniskill_pick_cube.yaml) — paired SimEnvironment config.
+- [`scenes/benchmarks/maniskill3_pick_cube.yaml`](../../scenes/benchmarks/maniskill3_pick_cube.yaml) — bare-backend MS3 PickCube wiring test.
 - [CLAUDE.md §6.4](../../CLAUDE.md) — rSkill packaging contract.

eval/maniskill3_franka_pick_cube.json

@@ -0,0 +1,50 @@
+{
+  "schema_version": "0.1",
+  "source": {
+    "paper": "https://arxiv.org/abs/2410.00425",
+    "arxiv": "https://arxiv.org/abs/2410.00425",
+    "model_variant": "smolvla",
+    "evaluated_by": "OpenRAL:openral benchmark run",
+    "reproduced_locally": true,
+    "reproduction_planned": null,
+    "reproduction_cli": "openral benchmark run --suite maniskill3_franka_pick_cube --rskill rskill://smolvla-maniskill-franka",
+    "table": null,
+    "status": "reproduced"
+  },
+  "benchmark": {
+    "name": "ManiSkill3 PickCube (Franka Panda)",
+    "dataset": null,
+    "protocol": "10 episodes per task, success_key=success, max_steps=100",
+    "robot": "franka_panda",
+    "simulator": "ManiSkill3 (SAPIEN, GPU)"
+  },
+  "eval_config": {
+    "n_episodes": 10,
+    "seeds": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      5,
+      6,
+      7,
+      8,
+      9
+    ],
+    "success_key": "success",
+    "max_steps": 100,
+    "vla_id": "smolvla",
+    "weights_uri": "rskill://smolvla-maniskill-franka"
+  },
+  "results": {
+    "maniskill3/PickCube-v1_success_rate": 0.1,
+    "avg_success_rate": 0.1,
+    "n_tasks": 1,
+    "n_episodes_per_task": 10,
+    "n_episodes_total": 10,
+    "mean_step_latency_ms_avg": 23.350579773366917
+  },
+  "baselines": {},
+  "trace_id": null
+}

rskill.yaml

@@ -40,11 +40,12 @@
 # processors.
 
 # ── Identity ───────────────────────────────────────────────────────────────
-schema_version: "1"
-name: "AdrianLlopart/rskill-smolvla-maniskill-franka"
+schema_version: "0.1"
+name: "OpenRAL/rskill-smolvla-maniskill-franka"
 version: "0.1.0"
 license: "apache-2.0"
 role: "s1"
+kind: "vla"  # ADR-00XX: rSkill kind discriminator. "vla" = learnable Vision-Language-Action policy.
 
 # ── Policy identity ────────────────────────────────────────────────────────
 model_family: "smolvla"
@@ -116,7 +117,7 @@ latency_budget:
 
 # ── Provenance ─────────────────────────────────────────────────────────────
 # No benchmarks shipped yet — see eval/.gitkeep. The headline LiftCube
-# success rate will be populated by `ral benchmark run` once the MS3
+# success rate will be populated by `openral benchmark run` once the MS3
 # backend supports the wrist-camera surface (see header comment).
 paper_url: "https://arxiv.org/abs/2506.01844"
 dataset_uri: "hf://Calvert0921/SmolVLA_LiftCube_Franka_1000"
@@ -128,9 +129,9 @@ description: >
   Panda lifting a cube in ManiSkill3 SAPIEN). Action chunks of length
   50 across overhead + wrist RGB views and a 9-D Franka qpos state.
   Runs end-to-end on ManiSkill3 PickCube-v1 with a live SAPIEN viewer
-  via `ral sim run --view`.
+  via `openral sim run --view`.
 
-# ADR-0021 — action vocabulary surfaced to the reasoner LLM tool palette.
+# ADR-0022 — action vocabulary surfaced to the reasoner LLM tool palette.
 # "lift" is not in the RSkillAction enum; the lift motion is expressed
 # as a pick (the post-grasp upward motion is implicit in the policy's
 # closed-loop control).
@@ -146,3 +147,6 @@ scenes:
 # to bind the LeRobot v3 `action` feature shape).
 action_contract:
   dim: 8
+
+benchmarks:
+  maniskill3_franka_pick_cube: 0.1