Duplicate from AdrianLlopart/rskill-act-aloha

.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 +1,101 @@
+---
+tags:
+  - OpenRAL
+  - rskill
+  - act
+  - lerobot
+  - aloha
+  - bimanual
+  - manipulation
+license: mit
+language:
+  - en
+---
+
+# rskill-act-aloha
+
+> **OpenRAL rSkill** — ACT (Action Chunking Transformer) finetuned on
+> the ALOHA bimanual cube-transfer task, packaged for `OpenRAL`.
+
+This package wraps
+[`lerobot/act_aloha_sim_transfer_cube_human`](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human)
+with a `rskill.yaml` manifest that adds capability checking, license
+surfacing, latency budgets, and local registry integration. It does
+**not** copy model weights.
+
+## Upstream model
+
+| Field | Value |
+| --- | --- |
+| Source repo | [`lerobot/act_aloha_sim_transfer_cube_human`](https://huggingface.co/lerobot/act_aloha_sim_transfer_cube_human) |
+| Paper | [arxiv:2304.13705](https://arxiv.org/abs/2304.13705) — *Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware* (Zhao et al., 2023) |
+| License | MIT |
+| Parameters | ~52 M (transformer encoder-decoder) |
+| Action chunk | 100 |
+| Benchmark | ALOHA bimanual cube-transfer (`gym-aloha`) |
+
+> **Note.** The published checkpoint predates lerobot's
+> `PolicyProcessorPipeline` migration and ships **without normalisation
+> buffers**. See `tests/sim/test_aloha_bimanual_act_aloha.py` for the resulting
+> numerical-contract caveats.
+
+## Supported robots
+
+| Robot | Embodiment tag | Status | Notes |
+| --- | --- | --- | --- |
+| ALOHA bimanual (Trossen) — `gym-aloha` MuJoCo | `aloha`, `lerobot` | ✓ sim | 14-DoF (2 × 7-DoF arms with parallel grippers) |
+
+## Sensors required
+
+| Key | Type | Resolution | Format |
+| --- | --- | --- | --- |
+| `observation.images.top` | RGB camera | 640 × 480 | `float32` |
+
+ACT for ALOHA cube-transfer ships with a single top-down RGB stream. No
+wrist or third-person view.
+
+## Manifest summary
+
+| Field | Value |
+| --- | --- |
+| `name` | `AdrianLlopart/rskill-act-aloha` |
+| `version` | `0.1.0` |
+| `license` | `mit` |
+| `role` | `s1` |
+| `embodiment_tags` | `aloha`, `lerobot` |
+| `runtime` / `quantization.dtype` | `pytorch` / `fp32` |
+| `weights_uri` | `hf://lerobot/act_aloha_sim_transfer_cube_human` |
+| `latency_budget.per_chunk_ms` | 25 ms (warm; bf16 autocast ≈ 12 ms on RTX 4070 Laptop) |
+| `latency_budget.warmup_ms` | 5 000 ms |
+| `latency_budget.load_ms` | 10 000 ms |
+| `commercial_use_allowed` | `true` |
+
+Full schema: `openral_core.RSkillManifest` —
+`python/core/src/openral_core/schemas.py`.
+
+## Reproduction
+
+```bash
+git clone https://github.com/AdrianLlopart/openral && cd OpenRAL
+just bootstrap && uv sync --all-packages --group sim
+
+# End-to-end via the canonical SimEnvironment config:
+just sim-act-aloha
+# which runs:
+#     ral sim run --config examples/sim/act_aloha_transfer_cube.yaml --save-video
+
+# Sim test (real gym-aloha MuJoCo with contact dynamics):
+uv run pytest tests/sim/test_aloha_bimanual_act_aloha.py -v -m sim
+```
+
+## License
+
+This rSkill package (`rskill.yaml`, `README.md`) is **MIT** to match the
+upstream weights. Commercial use is allowed
+(`commercial_use_allowed: true`).
+
+## See also
+
+- [`robots/aloha_bimanual/README.md`](../../robots/aloha_bimanual/README.md) — RobotDescription manifest.
+- [`examples/sim/act_aloha_transfer_cube.yaml`](../../examples/sim/act_aloha_transfer_cube.yaml) — paired SimEnvironment config.
+- [`docs/reference/vla_compatibility.md`](../../docs/reference/vla_compatibility.md) — VLA × Robot × Sim matrix.

eval/README.md

@@ -0,0 +1,14 @@
+# `rskills/act-aloha/eval/` — benchmark results
+
+`aloha_transfer_cube.json` is the ALOHA bimanual cube-transfer benchmark
+result block for this rSkill. Validated against
+[`openral_core.RSkillEvalResult`](../../../docs/reference/schemas/RSkillEvalResult.json)
+at load time by the `rSkill` loader and surfaced by `ral benchmark report`.
+
+| Field | Value |
+| --- | --- |
+| Source | Zhao et al., 2023 — *Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware* (arxiv:2304.13705) |
+| Benchmark | ALOHA bimanual cube transfer (`gym_aloha/AlohaTransferCube-v0`) |
+| Robot | Trossen ALOHA (2 × 7-DoF + parallel grippers, 14-DoF action) |
+| Reproduced locally? | ✗ — paper-only. `tests/sim/test_aloha_bimanual_act_aloha.py` runs a single episode for IO + latency verification but does not aggregate the 50-trial protocol. |
+| Reproduce | `just sim-act-aloha` (single episode); raise `--n-episodes 50` for the full paper protocol. |

eval/aloha_transfer_cube.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_transfer_cube --rskill rskill://act-aloha",
+    "table": null,
+    "status": "reproduced"
+  },
+  "benchmark": {
+    "name": "ALOHA bimanual cube transfer (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"
+  },
+  "results": {
+    "aloha_transfer_cube/0_success_rate": 0.82,
+    "avg_success_rate": 0.82,
+    "n_tasks": 1,
+    "n_episodes_per_task": 50,
+    "n_episodes_total": 50,
+    "mean_step_latency_ms_avg": 10.794657473535338
+  },
+  "baselines": {}
+}

rskill.yaml

@@ -0,0 +1,74 @@
+# rSkill manifest — OpenRAL packaging format V1 (CLAUDE.md §6.4)
+# Wraps: lerobot/act_aloha_sim_transfer_cube_human (MIT)
+# Paper: Zhao et al., 2023 — Action Chunking Transformer.
+
+schema_version: "1"
+
+name: "AdrianLlopart/rskill-act-aloha"
+version: "0.1.0"
+license: "mit"
+role: "s1"
+
+model_family: "act"
+
+# Bimanual ALOHA (2 × 7-DoF arms = 14-DoF action space). Used by
+# tests/sim/test_aloha_bimanual_act_aloha.py (gym-aloha MuJoCo).
+embodiment_tags:
+  - "aloha"
+
+capabilities_required: {}
+
+# ACT for ALOHA cube-transfer ships with a single top-down 480×640 RGB stream.
+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_transfer_cube_human"
+
+chunk_size: 100
+
+latency_budget:
+  # Reference-host measurement (RTX 4070 Laptop, CUDA 12.8, PyTorch 2.10)
+  # of the warm full-chunk inference is 16 ms; bf16 autocast is ~12 ms.
+  # We pin per_chunk_ms to 25 ms to keep the canonical
+  # "tolerance_pct=100 → 2× ceiling" pattern (giving a 50 ms test ceiling,
+  # matching the previous _WARM_CHUNK_CEILING_S = 0.050).
+  per_chunk_ms: 25.0
+  warmup_ms: 5000.0
+  load_ms: 10000.0
+
+fallback_skill_id: null
+
+# Headline success rate from skills/act-aloha/eval/aloha_transfer_cube.json
+# (50 episodes via `ral benchmark run`).
+benchmarks:
+  aloha_transfer_cube: 0.82
+
+# Policy-class identity; ACT manages its own preprocessing / state
+# contract inside the lerobot ACTPolicy so nothing else needs to move.
+policy_id: "act"
+
+paper_url: "https://arxiv.org/abs/2304.13705"
+source_repo: "hf://lerobot/act_aloha_sim_transfer_cube_human"
+
+description: >
+  Action Chunking Transformer (~52M-param encoder-decoder) finetuned on
+  the ALOHA bimanual cube-transfer demonstration set. Action chunks of
+  length 100. The published checkpoint predates lerobot's
+  PolicyProcessorPipeline migration and ships without normalisation
+  buffers — see tests/sim/test_aloha_bimanual_act_aloha.py for the
+  resulting numerical-contract caveats.