AdrianLlopart commited on
Commit
d4442f6
Β·
verified Β·
1 Parent(s): f19fbdd

chore: publish rSkill OpenRAL/rskill-act-aloha v0.1.0

Browse files
Files changed (4) hide show
  1. README.md +4 -4
  2. eval/README.md +1 -1
  3. eval/aloha_transfer_cube.json +3 -3
  4. rskill.yaml +49 -18
README.md CHANGED
@@ -58,7 +58,7 @@ wrist or third-person view.
58
 
59
  | Field | Value |
60
  | --- | --- |
61
- | `name` | `AdrianLlopart/rskill-act-aloha` |
62
  | `version` | `0.1.0` |
63
  | `license` | `mit` |
64
  | `role` | `s1` |
@@ -76,13 +76,13 @@ Full schema: `openral_core.RSkillManifest` β€”
76
  ## Reproduction
77
 
78
  ```bash
79
- git clone https://github.com/AdrianLlopart/openral && cd OpenRAL
80
  just bootstrap && uv sync --all-packages --group sim
81
 
82
  # End-to-end via the canonical SimEnvironment config:
83
  just sim-act-aloha
84
  # which runs:
85
- # ral sim run --config examples/sim/act_aloha_transfer_cube.yaml --save-video
86
 
87
  # Sim test (real gym-aloha MuJoCo with contact dynamics):
88
  uv run pytest tests/sim/test_aloha_bimanual_act_aloha.py -v -m sim
@@ -97,5 +97,5 @@ upstream weights. Commercial use is allowed
97
  ## See also
98
 
99
  - [`robots/aloha_bimanual/README.md`](../../robots/aloha_bimanual/README.md) β€” RobotDescription manifest.
100
- - [`examples/sim/act_aloha_transfer_cube.yaml`](../../examples/sim/act_aloha_transfer_cube.yaml) β€” paired SimEnvironment config.
101
  - [`docs/reference/vla_compatibility.md`](../../docs/reference/vla_compatibility.md) β€” VLA Γ— Robot Γ— Sim matrix.
 
58
 
59
  | Field | Value |
60
  | --- | --- |
61
+ | `name` | `OpenRAL/rskill-act-aloha` |
62
  | `version` | `0.1.0` |
63
  | `license` | `mit` |
64
  | `role` | `s1` |
 
76
  ## Reproduction
77
 
78
  ```bash
79
+ git clone https://github.com/OpenRAL/openral && cd OpenRAL
80
  just bootstrap && uv sync --all-packages --group sim
81
 
82
  # End-to-end via the canonical SimEnvironment config:
83
  just sim-act-aloha
84
  # which runs:
85
+ # openral sim run --config scenes/benchmarks/act_aloha_transfer_cube.yaml --save-video
86
 
87
  # Sim test (real gym-aloha MuJoCo with contact dynamics):
88
  uv run pytest tests/sim/test_aloha_bimanual_act_aloha.py -v -m sim
 
97
  ## See also
98
 
99
  - [`robots/aloha_bimanual/README.md`](../../robots/aloha_bimanual/README.md) β€” RobotDescription manifest.
100
+ - [`scenes/benchmarks/act_aloha_transfer_cube.yaml`](../../scenes/benchmarks/act_aloha_transfer_cube.yaml) β€” paired SimEnvironment config.
101
  - [`docs/reference/vla_compatibility.md`](../../docs/reference/vla_compatibility.md) β€” VLA Γ— Robot Γ— Sim matrix.
eval/README.md CHANGED
@@ -3,7 +3,7 @@
3
  `aloha_transfer_cube.json` is the ALOHA bimanual cube-transfer benchmark
4
  result block for this rSkill. Validated against
5
  [`openral_core.RSkillEvalResult`](../../../docs/reference/schemas/RSkillEvalResult.json)
6
- at load time by the `rSkill` loader and surfaced by `ral benchmark report`.
7
 
8
  | Field | Value |
9
  | --- | --- |
 
3
  `aloha_transfer_cube.json` is the ALOHA bimanual cube-transfer benchmark
4
  result block for this rSkill. Validated against
5
  [`openral_core.RSkillEvalResult`](../../../docs/reference/schemas/RSkillEvalResult.json)
6
+ at load time by the `rSkill` loader and surfaced by `openral benchmark report`.
7
 
8
  | Field | Value |
9
  | --- | --- |
eval/aloha_transfer_cube.json CHANGED
@@ -1,13 +1,13 @@
1
  {
2
- "schema_version": "1",
3
  "source": {
4
  "paper": "https://arxiv.org/abs/2304.13705",
5
  "arxiv": "https://arxiv.org/abs/2304.13705",
6
  "model_variant": "act",
7
- "evaluated_by": "OpenRAL:ral benchmark run",
8
  "reproduced_locally": true,
9
  "reproduction_planned": null,
10
- "reproduction_cli": "ral benchmark run --suite aloha_transfer_cube --rskill rskill://act-aloha",
11
  "table": null,
12
  "status": "reproduced"
13
  },
 
1
  {
2
+ "schema_version": "0.1",
3
  "source": {
4
  "paper": "https://arxiv.org/abs/2304.13705",
5
  "arxiv": "https://arxiv.org/abs/2304.13705",
6
  "model_variant": "act",
7
+ "evaluated_by": "OpenRAL:openral benchmark run",
8
  "reproduced_locally": true,
9
  "reproduction_planned": null,
10
+ "reproduction_cli": "openral benchmark run --suite aloha_transfer_cube --rskill rskill://act-aloha",
11
  "table": null,
12
  "status": "reproduced"
13
  },
rskill.yaml CHANGED
@@ -1,23 +1,32 @@
1
  # rSkill manifest β€” OpenRAL packaging format V1 (CLAUDE.md Β§6.4)
2
  # Wraps: lerobot/act_aloha_sim_transfer_cube_human (MIT)
3
  # Paper: Zhao et al., 2023 β€” Action Chunking Transformer.
4
-
5
- schema_version: "1"
6
-
7
- name: "AdrianLlopart/rskill-act-aloha"
 
 
 
 
 
 
 
 
8
  version: "0.1.0"
9
  license: "mit"
10
  role: "s1"
 
11
 
 
12
  model_family: "act"
13
 
 
14
  # Bimanual ALOHA (2 Γ— 7-DoF arms = 14-DoF action space). Used by
15
  # tests/sim/test_aloha_bimanual_act_aloha.py (gym-aloha MuJoCo).
16
  embodiment_tags:
17
  - "aloha"
18
 
19
- capabilities_required: {}
20
-
21
  # ACT for ALOHA cube-transfer ships with a single top-down 480Γ—640 RGB stream.
22
  sensors_required:
23
  - modality: "rgb"
@@ -30,17 +39,25 @@ sensors_required:
30
  # robots/aloha_bimanual/robot.yaml.
31
  actuators_required:
32
  - kind: "joint_position"
 
 
33
 
 
34
  runtime: "pytorch"
35
-
36
  quantization:
37
  dtype: "fp32"
38
  backend: "pytorch"
39
-
40
  weights_uri: "hf://lerobot/act_aloha_sim_transfer_cube_human"
41
 
42
- chunk_size: 100
 
 
 
43
 
 
 
 
 
44
  latency_budget:
45
  # Reference-host measurement (RTX 4070 Laptop, CUDA 12.8, PyTorch 2.10)
46
  # of the warm full-chunk inference is 16 ms; bf16 autocast is ~12 ms.
@@ -48,20 +65,13 @@ latency_budget:
48
  # "tolerance_pct=100 β†’ 2Γ— ceiling" pattern (giving a 50 ms test ceiling,
49
  # matching the previous _WARM_CHUNK_CEILING_S = 0.050).
50
  per_chunk_ms: 25.0
51
- warmup_ms: 5000.0
52
- load_ms: 10000.0
53
-
54
- fallback_skill_id: null
55
 
 
56
  # Headline success rate from skills/act-aloha/eval/aloha_transfer_cube.json
57
- # (50 episodes via `ral benchmark run`).
58
  benchmarks:
59
  aloha_transfer_cube: 0.82
60
 
61
- # Policy-class identity; ACT manages its own preprocessing / state
62
- # contract inside the lerobot ACTPolicy so nothing else needs to move.
63
- policy_id: "act"
64
-
65
  paper_url: "https://arxiv.org/abs/2304.13705"
66
  source_repo: "hf://lerobot/act_aloha_sim_transfer_cube_human"
67
 
@@ -72,3 +82,24 @@ description: >
72
  PolicyProcessorPipeline migration and ships without normalisation
73
  buffers β€” see tests/sim/test_aloha_bimanual_act_aloha.py for the
74
  resulting numerical-contract caveats.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  # rSkill manifest β€” OpenRAL packaging format V1 (CLAUDE.md Β§6.4)
2
  # Wraps: lerobot/act_aloha_sim_transfer_cube_human (MIT)
3
  # Paper: Zhao et al., 2023 β€” Action Chunking Transformer.
4
+ #
5
+ # LEGACY PROCESSOR PATH: this checkpoint pre-dates lerobot's
6
+ # PolicyProcessorPipeline migration and ships its norm stats inside
7
+ # model.safetensors. The schema's processors block is therefore omitted;
8
+ # the ACT adapter dispatches on manifest.processors is None and falls
9
+ # back to the snapshot_download + _try_load_act_norm_stats path. Migrating
10
+ # to per-file URIs would require re-publishing the upstream checkpoint
11
+ # and is tracked as a follow-up.
12
+
13
+ # ── Identity ───────────────────────────────────────────────────────────────
14
+ schema_version: "0.1"
15
+ name: "OpenRAL/rskill-act-aloha"
16
  version: "0.1.0"
17
  license: "mit"
18
  role: "s1"
19
+ kind: "vla" # ADR-00XX: rSkill kind discriminator. "vla" = learnable Vision-Language-Action policy.
20
 
21
+ # ── Policy identity ────────────────────────────────────────────────────────
22
  model_family: "act"
23
 
24
+ # ── Compatibility contract ─────────────────────────────────────────────────
25
  # Bimanual ALOHA (2 Γ— 7-DoF arms = 14-DoF action space). Used by
26
  # tests/sim/test_aloha_bimanual_act_aloha.py (gym-aloha MuJoCo).
27
  embodiment_tags:
28
  - "aloha"
29
 
 
 
30
  # ACT for ALOHA cube-transfer ships with a single top-down 480Γ—640 RGB stream.
31
  sensors_required:
32
  - modality: "rgb"
 
39
  # robots/aloha_bimanual/robot.yaml.
40
  actuators_required:
41
  - kind: "joint_position"
42
+ control_mode_semantics:
43
+ mode: "absolute"
44
 
45
+ # ── Runtime / weights ──────────────────────────────────────────────────────
46
  runtime: "pytorch"
 
47
  quantization:
48
  dtype: "fp32"
49
  backend: "pytorch"
 
50
  weights_uri: "hf://lerobot/act_aloha_sim_transfer_cube_human"
51
 
52
+ # ── Preprocessing (all knobs needed to interpret IO) ───────────────────────
53
+ # processors omitted β€” legacy path; norm stats live inside model.safetensors.
54
+ # ACT manages its own preprocessing / state contract inside the lerobot
55
+ # ACTPolicy so nothing else needs to move.
56
 
57
+ # ── Execution semantics ────────────────────────────────────────────────────
58
+ chunk_size: 100
59
+ # n_action_steps omitted β€” ACT default is 1 (per-step re-inference +
60
+ # temporal ensembling, paper-faithful).
61
  latency_budget:
62
  # Reference-host measurement (RTX 4070 Laptop, CUDA 12.8, PyTorch 2.10)
63
  # of the warm full-chunk inference is 16 ms; bf16 autocast is ~12 ms.
 
65
  # "tolerance_pct=100 β†’ 2Γ— ceiling" pattern (giving a 50 ms test ceiling,
66
  # matching the previous _WARM_CHUNK_CEILING_S = 0.050).
67
  per_chunk_ms: 25.0
 
 
 
 
68
 
69
+ # ── Provenance ─────────────────────────────────────────────────────────────
70
  # Headline success rate from skills/act-aloha/eval/aloha_transfer_cube.json
71
+ # (50 episodes via `openral benchmark run`).
72
  benchmarks:
73
  aloha_transfer_cube: 0.82
74
 
 
 
 
 
75
  paper_url: "https://arxiv.org/abs/2304.13705"
76
  source_repo: "hf://lerobot/act_aloha_sim_transfer_cube_human"
77
 
 
82
  PolicyProcessorPipeline migration and ships without normalisation
83
  buffers β€” see tests/sim/test_aloha_bimanual_act_aloha.py for the
84
  resulting numerical-contract caveats.
85
+
86
+ # ADR-0022 β€” action vocabulary surfaced to the reasoner LLM tool
87
+ # palette so it can pick this skill by what it does (action verb +
88
+ # object + scene), not just by its slug.
89
+ actions:
90
+ - "transfer"
91
+ - "pick"
92
+ - "place"
93
+ objects:
94
+ - "cube"
95
+ scenes:
96
+ - "tabletop"
97
+
98
+ # ADR-0019 β€” per-checkpoint action contract (consumed by the dataset bridge
99
+ # to bind the LeRobot v3 `action` feature shape).
100
+ action_contract:
101
+ dim: 14
102
+
103
+ # ADR-0019 β€” per-checkpoint state contract.
104
+ state_contract:
105
+ dim: 14