rskill-gr00t_n17-franka_panda-libero_spatial-bf16

OpenRAL rSkill — NVIDIA Isaac GR00T N1.7 (3B) finetuned on the LIBERO benchmark, packaged for the OpenRAL robot agent framework.

This package wraps nvidia/GR00T-N1.7-LIBERO with a rskill.yaml manifest that adds capability checking, license surfacing, latency budgets, and local registry integration. It does not copy model weights.

Runtime status. The manifest, license posture, and model_family: gr00t are wired and tested, and the runtime adapter has landed — openral_sim.policies.gr00t runs GR00T N1.7 in-process under the workspace's Python 3.12 via lerobot 0.6.0's native GrootPolicy (backbone-only NF4 quantization), the same pattern as the SmolVLA/π0.5 adapters — not the out-of-process sidecar originally planned; there is no tools/gr00t_sidecar.py. (tools/behavior_groot_sidecar.py is unrelated: it drives the separate, external Python 3.10 wensi-ai/Isaac-GR00T runtime for the official 2026 BEHAVIOR-1K organizer checkpoint, not this rSkill.) eval/libero_spatial.json carries reproduced_locally: true.

Upstream model, architecture & training

Field	Value
Source repo	nvidia/GR00T-N1.7-LIBERO
Base model	nvidia/GR00T-N1.7-3B
VLM backbone	Cosmos-Reason2-2B (SigLip2 vision encoder)
Paper	arXiv:2503.14734 — GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
Parameters	~3.1 B
License	NVIDIA Open Model License Agreement (commercial use permitted)
Pretraining	20K hours EgoScale human video + diverse robot demonstrations
Finetune	LIBERO task suite (Franka Panda, MuJoCo via robosuite)

GR00T is a cross-embodiment foundation model with variable-dimension proprioception and per-embodiment action heads. This checkpoint specializes the N1.7 base on the LIBERO Franka embodiment.

Supported robots / embodiments

Robot	Embodiment tag	Status	Notes
Franka Panda (LIBERO sim)	franka_panda	packaged	Native finetune embodiment; live eval in GR00T backend epic PR2

GR00T exposes a LIBERO_PANDA embodiment tag internally; OpenRAL maps it to the canonical franka_panda embodiment from robots/.

Sensors / observation contract

Key	Type	Min resolution	Description
observation.images.camera1	RGB camera	224 × 224	Agentview / overhead
observation.images.camera2	RGB camera	224 × 224	Wrist / end-effector
state	Proprioception	(8,)	End-effector pose + gripper, LIBERO layout

The policy emits a 16-step action chunk; each action is 7-D (delta_ee_6d_plus_gripper). Images and state are normalized inside the GR00T checkpoint's own experiment_cfg metadata rather than a lerobot processor pipeline — hence no processors block in the manifest.

Manifest summary

Field	Value
name	OpenRAL/rskill-gr00t_n17-franka_panda-libero_spatial-bf16
version	0.1.0
license	nvidia_open_model (commercial OK)
role	s1
model_family	gr00t
embodiment_tags	franka_panda
runtime	pytorch (out-of-process sidecar)
quantization.dtype	int4 (NF4 at load; stored bf16 → quantization.extra.stored_dtype)
weights_uri	hf://nvidia/GR00T-N1.7-LIBERO
chunk_size	16
state_contract.dim / action_contract.dim	8 / 7
latency_budget.per_chunk_ms	1500 ms (sidecar round-trip + 3B inference)

Full schema: openral_core.schemas.RSkillManifest.

Hardware

GR00T N1.7-3B (bf16, ~6 GB weights) plus the Cosmos-Reason VLM does not fit on an 8 GB GPU without NF4 quantization; the sidecar (GR00T backend epic PR2) follows the NF4 isolated-venv recipe used by the RLDX and detector sidecars. A ≥ 16 GB GPU can skip packing with OPENRAL_QUANTIZATION_DTYPE=fp32 (the unpacked load runs fp32 params; there is no bf16 path).

License

This rSkill package (rskill.yaml, README.md) is Apache-2.0.

The wrapped model weights (nvidia/GR00T-N1.7-LIBERO) are governed by the NVIDIA Open Model License Agreement, which permits commercial use. This is the key distinction from GR00T N1 / N1.5 / N1.6, which ship under the NVIDIA OneWay Noncommercial License and are blocked in commercial deployments by the OpenRAL loader unless OPENRAL_ALLOW_NONCOMMERCIAL=1 is set (CLAUDE.md §3).