Duplicate from AdrianLlopart/rskill-act-aloha-insertion

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,67 @@
+---
+tags:
+  - OpenRAL
+  - rskill
+  - act
+  - lerobot
+  - aloha
+  - bimanual
+  - manipulation
+  - insertion
+license: mit
+language:
+  - en
+---
+
+# rskill-act-aloha-insertion
+
+> **OpenRAL rSkill (custom example)** — ACT (Action Chunking Transformer)
+> finetuned on the ALOHA bimanual **peg-insertion** task, packaged for
+> `OpenRAL`.
+
+This package wraps
+[`lerobot/act_aloha_sim_insertion_human`](https://huggingface.co/lerobot/act_aloha_sim_insertion_human)
+with a `rskill.yaml` manifest that adds capability checking, license
+surfacing, latency budgets, and local registry integration. It does
+**not** copy model weights.
+
+It is the harder sibling of [`rskill-act-aloha`](../act-aloha) (cube
+transfer) and demonstrates how a single packaging format covers multiple
+task-specific checkpoints from the same paper. The runnable demo lives at
+`examples/sim/custom_act_aloha_insertion.yaml` and is wired into the
+top-level `just sim-custom` recipe.
+
+## Upstream model
+
+| Field | Value |
+| --- | --- |
+| Source repo | [`lerobot/act_aloha_sim_insertion_human`](https://huggingface.co/lerobot/act_aloha_sim_insertion_human) |
+| Architecture | Action Chunking Transformer (~52M params, chunk=100) |
+| Task | gym-aloha `AlohaInsertion-v0` (bimanual peg-in-socket) |
+| License | MIT |
+| Paper | Zhao et al., 2023 — *Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware* ([arXiv 2304.13705](https://arxiv.org/abs/2304.13705)) |
+
+## Why no `eval/` block?
+
+This skill is shipped as a **custom-example** package, not as a
+reproduced benchmark entry. The paper's headline number for sim ALOHA
+insertion is markedly lower than the cube-transfer figure (the task is
+harder and the upstream protocol uses different camera intrinsics). We
+deliberately omit `eval/` rather than copy paper numbers without an
+internal reproduction; per CLAUDE.md §6.4 that omission must be
+documented — this section is that documentation. Add `eval/aloha_insertion.json`
+once a local reproduction lands.
+
+## Run it
+
+```bash
+just sim-custom
+```
+
+…which is equivalent to:
+
+```bash
+MUJOCO_GL=egl uv run --group sim ral sim run \
+  --config examples/sim/custom_act_aloha_insertion.yaml \
+  --save-video example_videos
+```

eval/aloha_insertion.json

@@ -0,0 +1,89 @@
+{
+  "schema_version": "1",
+  "source": {
+    "paper": "https://arxiv.org/abs/2304.13705",
+    "arxiv": "https://arxiv.org/abs/2304.13705",
+    "model_variant": "act",
+    "evaluated_by": "OpenRAL:ral benchmark run",
+    "reproduced_locally": true,
+    "reproduction_planned": null,
+    "reproduction_cli": "ral benchmark run --suite aloha_insertion --rskill rskill://act-aloha-insertion",
+    "table": null,
+    "status": "reproduced"
+  },
+  "benchmark": {
+    "name": "ALOHA bimanual peg insertion (gym-aloha)",
+    "dataset": null,
+    "protocol": "50 episodes per task, success_key=is_success, max_steps=400",
+    "robot": "aloha_bimanual",
+    "simulator": "gym-aloha (MuJoCo)"
+  },
+  "eval_config": {
+    "n_episodes": 50,
+    "seeds": [
+      0,
+      1,
+      2,
+      3,
+      4,
+      5,
+      6,
+      7,
+      8,
+      9,
+      10,
+      11,
+      12,
+      13,
+      14,
+      15,
+      16,
+      17,
+      18,
+      19,
+      20,
+      21,
+      22,
+      23,
+      24,
+      25,
+      26,
+      27,
+      28,
+      29,
+      30,
+      31,
+      32,
+      33,
+      34,
+      35,
+      36,
+      37,
+      38,
+      39,
+      40,
+      41,
+      42,
+      43,
+      44,
+      45,
+      46,
+      47,
+      48,
+      49
+    ],
+    "success_key": "is_success",
+    "max_steps": 400,
+    "vla_id": "act",
+    "weights_uri": "rskill://rskills/act-aloha-insertion"
+  },
+  "results": {
+    "aloha_insertion/0_success_rate": 0.2,
+    "avg_success_rate": 0.2,
+    "n_tasks": 1,
+    "n_episodes_per_task": 50,
+    "n_episodes_total": 50,
+    "mean_step_latency_ms_avg": 10.16975808690558
+  },
+  "baselines": {}
+}

rskill.yaml

@@ -0,0 +1,71 @@
+# rSkill manifest — OpenRAL packaging format V1 (CLAUDE.md §6.4)
+# Wraps: lerobot/act_aloha_sim_insertion_human (MIT)
+# Paper: Zhao et al., 2023 — Action Chunking Transformer.
+#
+# Sibling of skills/act-aloha (which wraps the transfer-cube checkpoint).
+# This package targets the harder bimanual peg-in-socket *insertion* task
+# from the same paper. Both skills share the ALOHA 14-DoF action space and
+# the single 480x640 top camera contract.
+
+schema_version: "1"
+
+name: "AdrianLlopart/rskill-act-aloha-insertion"
+version: "0.1.0"
+license: "mit"
+role: "s1"
+
+model_family: "act"
+
+embodiment_tags:
+  - "aloha"
+
+capabilities_required: {}
+
+sensors_required:
+  - modality: "rgb"
+    vla_feature_key: "observation.images.top"
+    min_width: 640
+    min_height: 480
+
+# Output side (ADR-0013). For the canonical aloha bimanual embodiment the
+# loader auto-fills n_dof (14) + vla_action_key from
+# robots/aloha_bimanual/robot.yaml.
+actuators_required:
+  - kind: "joint_position"
+
+runtime: "pytorch"
+
+quantization:
+  dtype: "fp32"
+  backend: "pytorch"
+
+weights_uri: "hf://lerobot/act_aloha_sim_insertion_human"
+
+chunk_size: 100
+
+latency_budget:
+  # Same ACT architecture as the cube-transfer sibling; reference-host
+  # measurements match within noise. The 25 ms ceiling keeps the
+  # canonical tolerance_pct=100 → 50 ms test ceiling.
+  per_chunk_ms: 25.0
+  warmup_ms: 5000.0
+  load_ms: 10000.0
+
+fallback_skill_id: null
+
+# Headline success rate from skills/act-aloha-insertion/eval/aloha_insertion.json.
+benchmarks:
+  aloha_insertion: 0.20
+
+policy_id: "act"
+
+paper_url: "https://arxiv.org/abs/2304.13705"
+source_repo: "hf://lerobot/act_aloha_sim_insertion_human"
+
+description: >
+  ACT (~52M params, chunk=100) finetuned on the ALOHA bimanual
+  sim-insertion demonstration set. Like the transfer-cube sibling, the
+  published checkpoint predates lerobot's PolicyProcessorPipeline
+  migration and ships without normalisation buffers. Insertion is the
+  harder ALOHA task in the original paper; the 0.20 success rate
+  reflects that.