File size: 2,028 Bytes
c5ccfcf | 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 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 | # 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.
|