# rSkill manifest — OpenRAL packaging format V1 (CLAUDE.md §6.4)
# Wraps: Calvert0921/smolvla_franka_liftcube_1000
# Base:  lerobot/smolvla_base   (arxiv:2506.01844)
#
# Community SmolVLA finetune on a ManiSkill-style Franka LiftCube task,
# trained on the paired dataset Calvert0921/SmolVLA_LiftCube_Franka_1000
# (1000 demos). The upstream model card is empty so this rSkill is
# wired strictly from the checkpoint's `config.json`:
#
#   type: smolvla           → routed via the in-tree `smolvla` adapter.
#   input_features:
#     observation.state              shape (9,)        — Franka qpos
#                                                       (7 arm + 2 fingers)
#     observation.images.up          shape (3,256,256) — top / overhead view
#     observation.images.wrist       shape (3,256,256) — eye-in-hand view
#   output_features:
#     action                          shape (8,)        — 7 joint pos + 1 gripper
#   chunk_size: 50  n_action_steps: 50  flow-matching steps: 10
#
# Camera-name mapping
# -------------------
# The ManiSkill3 backend surfaces cameras under `camera1` / `camera2`
# (matching the franka_panda RobotDescription's `sensors` block); the
# `image_preprocessing.aliases` map below renames those to the model's
# `up` / `wrist` keys so the SmolVLA preprocessor finds them.
#
# End-to-end wiring
# -----------------
# The companion ManiSkill3 backend changes (this PR) surface every
# entry in `sensor_data` as `camera1` / `camera2` / ... in declaration
# order, plumb `backend_options.robot_uids` through to `gym.make` so
# the YAML's `panda_wristcam` brings in the wrist camera, and forward
# `task.max_steps` to MS3's `max_episode_steps` so the rollout isn't
# silently truncated at MS3's default 50 steps.
#
# The upstream model repo (Calvert0921/smolvla_franka_liftcube_1000)
# ships only config.json + model.safetensors + train_config.json — no
# `policy_*processor.json`. Rather than rely on the SmolVLA adapter's
# dataset-stats fallback at every load (a 404 + recompute against
# `dataset_uri`'s `meta/episodes_stats.jsonl`), the processor pair was
# rebuilt once via `make_pre_post_processors(policy.config,
# dataset_stats=...)` and uploaded to THIS rSkill's own HF repo
# (OpenRAL/rskill-smolvla-franka_panda-maniskill3-bf16). The rebuilt normalizer
# stats match the checkpoint's baked `normalize_inputs` /
# `normalize_targets` buffers exactly, so the processors are faithful to
# how the model was trained. The fallback path still exists for any
# other community finetune uploaded without processors.

# ── Identity ───────────────────────────────────────────────────────────────
schema_version: "0.1"
name: "OpenRAL/rskill-smolvla-franka_panda-maniskill3-bf16"
version: "0.1.0"
license: "apache-2.0"
role: "s1"
kind: "vla"  # rSkill kind discriminator. "vla" = learnable Vision-Language-Action policy.

# ── Policy identity ────────────────────────────────────────────────────────
model_family: "smolvla"

# ── Compatibility contract ─────────────────────────────────────────────────
# ManiSkill3 PickCube / LiftCube run on the Franka Panda agent — same
# embodiment as the LIBERO / RoboCasa Panda rSkills.
embodiment_tags:
  - "franka_panda"

# Two RGB camera streams matching the checkpoint's input_features. The
# franka_panda RobotDescription declares matching `camera1` (agentview /
# base_camera) and `camera2` (eye_in_hand / hand_camera) sensors so the
# compatibility check resolves cleanly.
sensors_required:
  - modality: "rgb"
    vla_feature_key: "observation.images.camera1"
    min_width: 256
    min_height: 256
  - modality: "rgb"
    vla_feature_key: "observation.images.camera2"
    min_width: 256
    min_height: 256

# Output side. 7 arm joints + 1 gripper command. For
# franka_panda the loader auto-fills n_dof + vla_action_key from
# robots/franka_panda/robot.yaml.
actuators_required:
  - kind: "joint_position"
    control_mode_semantics:
      mode: "absolute"

# ── Runtime / weights ──────────────────────────────────────────────────────
runtime: "pytorch"
quantization:
  dtype: "bf16"
  backend: "pytorch"
weights_uri: "hf://Calvert0921/smolvla_franka_liftcube_1000"

# ── Preprocessing (all knobs needed to interpret IO) ───────────────────────
# Per-file URIs for the lerobot PolicyProcessorPipeline. The upstream
# weights repo ships no `policy_*processor.json`, so the pair (plus the
# normalizer/unnormalizer `*.safetensors` state files referenced by the
# JSON `steps`) was rebuilt from the checkpoint config + training-dataset
# stats and uploaded to this rSkill's own HF repo. `materialize_processor_dir`
# downloads exactly these files (and their `state_file` siblings) — no 404,
# no dataset-stats recompute at load time.
processors:
  preprocessor_uri: "hf://OpenRAL/rskill-smolvla-franka_panda-maniskill3-bf16/policy_preprocessor.json"
  postprocessor_uri: "hf://OpenRAL/rskill-smolvla-franka_panda-maniskill3-bf16/policy_postprocessor.json"
# Per-checkpoint knobs the processor JSONs don't encode. The model was
# trained on raw SAPIEN renders (no 180° flip); cameras need an alias
# remap from the in-tree scene-side keys (camera1 / camera2) to the
# checkpoint's `up` / `wrist`.
image_preprocessing:
  flip_180: false
  aliases:
    top: "up"
    wrist: "wrist"
state_contract:
  dim: 9

# ── Execution semantics ────────────────────────────────────────────────────
# Matches the checkpoint's config.json (chunk_size=50, n_action_steps=50).
chunk_size: 50
n_action_steps: 50
latency_budget:
  per_chunk_ms: 200.0   # SmolVLA bf16 on a desktop GPU ≈ 80–150 ms; SAPIEN
                        # step adds another ~10–20 ms — generous headroom

# ── Provenance ─────────────────────────────────────────────────────────────
# No benchmarks shipped yet — see eval/.gitkeep. The headline LiftCube
# success rate will be populated by `openral benchmark run` once the MS3
# backend supports the wrist-camera surface (see header comment).
paper_url: "https://arxiv.org/abs/2506.01844"
dataset_uri: "hf://Calvert0921/SmolVLA_LiftCube_Franka_1000"
source_repo: "hf://Calvert0921/smolvla_franka_liftcube_1000"
# Task-data gate: this checkpoint was trained on the *LiftCube*
# dataset (a ManiSkill2-style "lift the cube to a height" task) which has NO
# equivalent env in ManiSkill3 3.0.1 — MS3 ships PickCube-v1 (grasp + place at
# a goal + stay static), a different task. Declaring the (MS3-absent) true task
# makes the benchmark runner refuse every real MS3 scene with
# ROSCapabilityMismatch. Verified empirically 2026-06-19: retargeted to
# PickCube-v1 it runs (50 steps, ~15 ms/step real inference) and attempts the
# cube but cannot satisfy PickCube's goal-placement success (success=False) —
# exactly the plausible-but-unsolvable rollout the gate exists to block. This
# rSkill therefore has no valid MS3 benchmark; replace with a PickCube-trained
# policy (e.g. the author's PegInsertion checkpoint -> PegInsertionSide-v1, or
# an MS3 OpenVLA-OFT policy) before pairing it with a maniskill scene.
evaluated_tasks: ["maniskill3/LiftCube"]

description: >
  SmolVLA (0.45 B, lerobot/smolvla_base) finetuned on
  Calvert0921/SmolVLA_LiftCube_Franka_1000 (1000 demos of a Franka
  Panda lifting a cube in ManiSkill3 SAPIEN). Action chunks of length
  50 across overhead + wrist RGB views and a 9-D Franka qpos state.
  Runs end-to-end on ManiSkill3 PickCube-v1 with a live SAPIEN viewer
  via `openral sim run --view`.

# Action vocabulary surfaced to the reasoner LLM tool palette.
# "lift" is not in the RSkillAction enum; the lift motion is expressed
# as a pick (the post-grasp upward motion is implicit in the policy's
# closed-loop control).
actions:
  - "pick"
  - "grasp"
objects:
  - "cube"
scenes:
  - "tabletop"

# Per-checkpoint action contract (consumed by the dataset bridge
# to bind the LeRobot v3 `action` feature shape).
action_contract:
  dim: 8

benchmarks:
  # PickCube-v1 via the curated franka_panda suite `maniskill3_panda`
  # (auto-filters to this rSkill's task). 0.0 = the shipped LiftCube
  # checkpoint cannot satisfy PickCube's goal-placement success (task
  # mismatch — see the MS3 LiftCube-gap note above); replace with a
  # PickCube-trained policy to lift this.
  maniskill3_panda: 0.0