rskill-gr00t_n17-r1pro-turning_on_radio-bf16

OpenRAL/rskill-gr00t_n17-r1pro-turning_on_radio-bf16

rskill-gr00t_n17-r1pro-turning_on_radio-bf16 is a robotics model from OpenRAL. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as other.

Requires OpenRAL Pro. This rSkill runs at int4 (NF4) (sub-16-bit weights). Under ADR-0123, open-source OpenRAL loads only fp32, fp16 and bf16 weights; quantized formats (int8, fp8, int4/NF4, nvfp4) need the OpenRAL Pr…

Downloads · 30 days

Access

Public

Updated Nov 7, 2026

Repo size

—

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.md11.3 KB · 62%

At a glance

Task: Robotics
License: other
Access: Public
Created: Aug 7, 2026
Updated: Nov 7, 2026
SHA: ed3f43af

Base models

nvidia/GR00T-N1.7-3B

Datasets

behavior-1k/2026-challenge-demos

Model card

From the Hugging Face model README

rskill-gr00t-n17-b1k-turning-on-radio

Requires OpenRAL Pro. This rSkill runs at int4 (NF4) (sub-16-bit weights). Under ADR-0123, open-source OpenRAL loads only fp32, fp16 and bf16 weights; quantized formats (int8, fp8, int4/NF4, nvfp4) need the OpenRAL Pro quantization package.

OpenRAL rSkill — the official 2026 BEHAVIOR-1K GR00T N1.7 turning_on_radio checkpoint for the simulated Galaxea R1 Pro.

This package contains the OpenRAL manifest and adapter configuration only. The organizer checkpoint is downloaded separately from the BEHAVIOR baseline page; weights are not copied into this repository.

Preview

The official task demonstration is available in the BEHAVIOR challenge gallery. No local rollout image is shipped until the checkpoint has been reproduced on this host.

What this skill does

The policy navigates the R1 Pro to a household radio and manipulates its controls to turn it on. It is task-specific to turning_on_radio.

Field	Value
Actions	rotate, push
Objects	radio, dial, button
Scenes	household living room
Embodiment	BEHAVIOR R1 Pro (`r1pro`)

How it works

The adapter runs the pinned wensi-ai/Isaac-GR00T behavior branch in an isolated Python 3.10 sidecar. The sidecar uses the upstream Gr00tPolicy and B1KPolicyWrapper unchanged, including temporal ensembling and the official R1Pro modality split. Two 8 GB-host memory measures are applied at load time: whole-model NF4 quantization of large linears, and replacing the Qwen3-VL lm_head with Identity (the wrapper consumes only hidden states, so the full-vocab logits projection is dead weight).

Observation -> action contract

Direction	Key	Shape	Notes
in	head RGB	HWC uint8	Official ZED head camera
in	left wrist RGB	HWC uint8	Left RealSense
in	right wrist RGB	HWC uint8	Right RealSense
in	proprioception	`(61,)` float32	Official `PROPRIOCEPTION_INDICES["R1Pro"]` order
out	action	`(23,)` float32	Base velocity 3 + torso 4 + arms 7+7 + grippers 1+1

The organizer wrapper converts the 61-D vector into base, torso, arm, and gripper state groups. It emits a 16-step GR00T action horizon and returns one temporally-ensembled 23-D action per evaluator step.

Upstream model / training

Field	Value
Source code	`wensi-ai/Isaac-GR00T@ace36d9`
Checkpoint	Organizer-provided `turning_on_radio` Google Drive checkpoint
Base model	`nvidia/GR00T-N1.7-3B`
Dataset	`behavior-1k/2026-challenge-demos`
Paper	arXiv:2503.14734
Parameters	approximately 3.1 B
Weights license	NVIDIA Open Model License, as the organizers' pinned runtime README states for its model weights

Supported robots

Robot	Embodiment tag	Status	Notes
Simulated Galaxea R1 Pro	`r1pro`	evaluator + deploy sim	Official BEHAVIOR observation/action contract

openral deploy sim uses robots/r1pro/robot.yaml, publishes the simulator's native 61-D policy state through WorldState, validates every typed action slot through the safety kernel, then atomically commits all six slots as one 23-D OmniGibson step.

Sensors required

Key	Modality	Min resolution	Format
`observation.images.head`	RGB	224 x 224	HWC uint8
`observation.images.left_wrist`	RGB	224 x 224	HWC uint8
`observation.images.right_wrist`	RGB	224 x 224	HWC uint8
`observation.state`	proprioception	`(61,)`	float32

Manifest summary

Field	Value
`name`	`OpenRAL/rskill-gr00t_n17-r1pro-turning_on_radio-bf16`
`version`	`0.1.0`
`license`	`nvidia_open_model`
`role`	`s1`
`model_family`	`gr00t_b1k`
`quantization.dtype`	`int4` (NF4 at load; stored `bf16` → `quantization.extra.stored_dtype`)
`runtime`	`pytorch` (the adapter drives an external Python 3.10 Isaac-GR00T sidecar), whole-model NF4
`weights_uri`	`local://checkpoints/behavior-groot-turning-on-radio`
`chunk_size`	16
`state_contract.dim` / `action_contract.dim`	61 / 23
`latency_budget.per_chunk_ms`	1500

Quick start

git clone https://github.com/wensi-ai/Isaac-GR00T \
  ~/.cache/openral/behavior-groot/source
git -C ~/.cache/openral/behavior-groot/source checkout \
  ace36d935b376fbf25cd56371e23877b95407c40
cd ~/.cache/openral/behavior-groot/source
uv sync --frozen --python 3.10

export OPENRAL_BEHAVIOR_GROOT_SIDECAR_PYTHON="$PWD/.venv/bin/python"
export OPENRAL_BEHAVIOR_GROOT_CHECKPOINT=/absolute/path/to/checkpoint

cd /path/to/openral
just sync --group behavior-groot
openral behavior serve \
  --rskill rskills/gr00t-n17-b1k-turning-on-radio \
  --task turning_on_radio

Full deploy graph:

openral deploy sim \
  --config scenes/deploy/behavior_r1pro.yaml \
  --initial-task "turn on the radio"

In the BEHAVIOR environment:

python -m omnigibson.eval.eval \
  --task-name turning_on_radio \
  --host 127.0.0.1 --port 8000 \
  --instance-indices 0 --num-rollouts 1 \
  --output-dir outputs/openral --write-video

Reproduction

The command above is the canonical one-rollout reproduction. Use public instances 0-9 for challenge reporting and keep the evaluator-generated JSON and videos unmodified.

Evaluation

eval/behavior.json records one OpenRAL episode (openral benchmark run --suite behavior --rskill rskills/gr00t-n17-b1k-turning-on-radio, seed 0, max 3224 steps): avg_success_rate 0.0. The full official-evaluator loop has also been reproduced locally on an 8 GB RTX 4070 Laptop GPU (NF4 sidecar 2.77 GiB inference peak alongside OmniGibson, ~0.8-1.3 steps/s): public instance 0 (ID 301), 3225 steps to timeout, success=false, q_score 0.0. Whether the zero q_score reflects NF4 degradation or the checkpoint's zero-shot behavior on this instance has not been isolated; no success-rate claim is made.

License

The OpenRAL adapter, manifest, and documentation are Apache-2.0. The organizer-provided fine-tuned checkpoint is under the NVIDIA Open Model License, which permits commercial use. The Google Drive artifact carries no license file of its own, but the organizers' pinned runtime, wensi-ai/Isaac-GR00T@ace36d9, states that its model weights are released under that license, and the challenge pages name no other terms.