schema_version: "0.1"
name: "OpenRAL/rskill-gr00t_n17-franka_panda-libero_spatial-bf16"
evaluated_tasks: ["libero_spatial"]  # benchmark task this checkpoint is validated for (gate)
version: "0.1.0"
# NVIDIA Open Model License — GR00T N1.7 permits commercial use, unlike the
# OneWay Noncommercial License on N1 / N1.5 / N1.6. The loader
# therefore does NOT block this skill in a commercial deployment.
license: "nvidia_open_model"
role: "s1"
kind: "vla"

# In-process GR00T N1.7 (following a later amendment to the GR00T backend
# design). lerobot 0.6.0 ships a native
# `GrootPolicy` (Cosmos-Reason2 / Qwen3-VL backbone) that loads under the
# workspace's Python 3.12, so the openral `gr00t` adapter now runs the 3B model
# in-process (NF4 backbone rewrite, ~5.2 GiB peak) instead of the retired
# Python-3.10 ZMQ sidecar. The `rldx` adapter still runs RLDX-1 (a GR00T-N1.5
# finetune native lerobot rejects) out-of-process. See
# openral_sim.policies.gr00t.
model_family: "gr00t"

embodiment_tags:
  - "franka_panda"

# GR00T uses its own embodiment namespace: the LIBERO checkpoint's tag is
# `libero_sim` (id 2 in embodiment_id.json). This selects the image+wrist_image
# views, the 8-D x,y,z,roll,pitch,yaw,gripper state / 7-D action modality, the
# gripper-sign flip (action_decode_transform=auto→libero), and the 8-of-16
# execution horizon. The adapter defaults to this; pinning it here keeps the
# rollout trace explicit. Override with vla.extra.embodiment_tag /
# OPENRAL_GR00T_EMBODIMENT_TAG for other GR00T checkpoints.
policy_extras:
  embodiment_tag: "libero_sim"

# LIBERO convention: agentview + wrist RGB. GR00T resizes internally; these
# are minimum source resolutions.
sensors_required:
  - modality: "rgb"
    vla_feature_key: "observation.images.camera1"
    min_width: 224
    min_height: 224
  - modality: "rgb"
    vla_feature_key: "observation.images.camera2"
    min_width: 224
    min_height: 224

actuators_required:
  - kind: "joint_position"
    control_mode_semantics:
      mode: "absolute"

runtime: "pytorch"
quantization:
  dtype: "bf16"
  backend: "pytorch"

# Root-level repackage of nvidia/GR00T-N1.7-LIBERO's `libero_spatial/` per-suite
# checkpoint (the upstream repo has NO root model.safetensors — its weights live
# in a subfolder amid DeepSpeed optimizer shards, so the bare repo id resolves to
# nothing loadable). This OpenRAL repo bundles the inference checkpoint at the
# repo root (sharded model-0000X-of-00002.safetensors + experiment_cfg +
# statistics.json + processor_config.json) so lerobot's `GrootPolicy.from_pretrained`
# loads it directly. The gr00t adapter then NF4-quantizes the Qwen3-VL backbone
# in place on load (~5.2 GiB peak VRAM; OPENRAL_GR00T_QUANTIZATION default nf4).
weights_uri: "hf://OpenRAL/rskill-gr00t_n17-franka_panda-libero_spatial-bf16"

# GR00T checkpoints carry their own normalization metadata (experiment_cfg +
# processor_config.json + statistics.json in the checkpoint) which lerobot's
# make_groot_pre_post_processors_from_pretrained reads directly, rather than
# lerobot PolicyProcessorPipeline JSONs, so this family declares no `processors`
# block.

# Upstream GR00T's own LIBERO eval env rotates BOTH cameras 180° before the
# policy — see Isaac-GR00T `gr00t/eval/sim/LIBERO/libero_env.py`:
#   "video.image":       obs["agentview_image"][::-1, ::-1]
#   "video.wrist_image": obs["robot0_eye_in_hand_image"][::-1, ::-1]
# so the checkpoint was trained on 180°-flipped frames (the standard LIBERO
# convention shared by every other libero rSkill: smolvla / pi05 / act / xvla /
# rldx1-ft-libero all set flip_180: true). Our lerobot LiberoEnv passes raw
# robosuite frames, so the adapter must apply the flip. Without this the policy
# saw upside-down images and scored 0/10 on libero_spatial (GPU-verified before
# the fix); the in-process gr00t adapter applies it in `_build_batch` off this flag.
image_preprocessing:
  flip_180: true

state_contract:
  dim: 8

chunk_size: 16
latency_budget:
  per_chunk_ms: 1500.0

paper_url: "https://arxiv.org/abs/2503.14734"
source_repo: "hf://nvidia/GR00T-N1.7-LIBERO"  # upstream provenance (libero_spatial/ subfolder)

description: >
  NVIDIA Isaac GR00T N1.7 (3B, Cosmos-Reason2-2B VLM backbone) finetuned on
  the LIBERO benchmark, packaged for OpenRAL. 7-D LIBERO action space
  (delta end-effector 6-DoF + gripper) over two RGB views. Runs in-process via
  lerobot 0.6.0's native GrootPolicy with an NF4-quantized backbone (~5.2 GiB
  peak, fits an 8 GB GPU). Open Model License — commercial use
  permitted.

actions:
  - "pick"
  - "place"
  - "open"
  - "close"
objects:
  - "bowl"
  - "cup"
  - "drawer"
  - "object"
scenes:
  - "tabletop"

action_contract:
  dim: 7
  representation: "delta_ee_6d_plus_gripper"

benchmarks:
  libero_spatial: 0.9