rskill.yaml

# 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.
#
# LEGACY PROCESSOR PATH: same shape as rskills/act-aloha — the mean/std
# norm stats live inside model.safetensors and the schema processors block
# is omitted. Modern lerobot ACTPolicy no longer hosts normalize_inputs/
# unnormalize_outputs modules, so it drops those keys on load ("Unexpected
# key(s)") — the ACT adapter re-reads them from the safetensors and applies
# them manually. Without that the policy runs unnormalised and flails.

# ── Identity ───────────────────────────────────────────────────────────────
schema_version: "0.1"
name: "OpenRAL/rskill-act-aloha-aloha_insertion-fp32"
version: "0.1.0"
license: "mit"
role: "s1"
kind: "vla"  # rSkill kind discriminator. "vla" = learnable Vision-Language-Action policy.

# ── Policy identity ────────────────────────────────────────────────────────
model_family: "act"

# ── Compatibility contract ─────────────────────────────────────────────────
embodiment_tags:
  - "aloha"

sensors_required:
  - modality: "rgb"
    vla_feature_key: "observation.images.top"
    min_width: 640
    min_height: 480

# Output side. 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"
    control_mode_semantics:
      mode: "absolute"

# ── Runtime / weights ──────────────────────────────────────────────────────
runtime: "pytorch"
quantization:
  dtype: "fp32"
  backend: "pytorch"
weights_uri: "hf://lerobot/act_aloha_sim_insertion_human"

# ── Preprocessing (all knobs needed to interpret IO) ───────────────────────
# processors omitted — legacy path; norm stats live inside model.safetensors.

# ── Execution semantics ────────────────────────────────────────────────────
chunk_size: 100
# n_action_steps omitted — ACT default is 1 (per-step re-inference +
# temporal ensembling, paper-faithful).
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

# ── Provenance ─────────────────────────────────────────────────────────────
# Headline success rate from skills/act-aloha-insertion/eval/aloha_insertion.json.
benchmarks:
  aloha_insertion: 0.0

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. Insertion is the harder ALOHA task
  in the original paper; the 0.20 success rate reflects that. See the
  norm-stats note above re: the legacy safetensors-resident buffers.

# Task-data gate: trained + validated on the ALOHA sim
# insertion task, so the benchmark runner accepts the aloha_insertion
# scene and refuses task-mismatched scenes with ROSCapabilityMismatch.
evaluated_tasks: ["aloha_insertion"]

# Action vocabulary surfaced to the reasoner LLM tool
# palette so it can pick this skill by what it does (action verb +
# object + scene), not just by its slug.
actions:
  - "insert"
  - "pick"
  - "place"
objects:
  - "peg"
  - "socket"
scenes:
  - "tabletop"

# Per-checkpoint action contract (consumed by the dataset bridge
# to bind the LeRobot v3 `action` feature shape).
action_contract:
  dim: 14
  # ACT on ALOHA emits absolute joint positions (2×(6 arm + 1 gripper)).
  representation: "joint_positions"
  # EXPLICIT joint units — the ALOHA sim (gym-aloha, MuJoCo qpos) records in
  # RADIANS (verified: observation.state normalizer mean∈[-0.22, 0.74],
  # std∈[0.10, 0.32] — every channel well under π). No deg↔rad conversion at
  # the policy boundary. Declared so the runner does not fall back to the
  # stats-magnitude heuristic.
  joint_units: "radians"

# Per-checkpoint state contract.
state_contract:
  dim: 14