OpenRAL — Runtime for VLA Robot Agents

⬛ GitHub 📄 Docs 💬 Discord Apache-2.0 ROS 2 Jazzy Python 3.12 16 robots · 20 rSkills · 22 benchmarks

A VLA model alone can't run on your robot. It needs a camera pipeline, an observation normaliser, an action de-normaliser, a safety wrapper, a replanning layer when it fails, and a way to log everything for later fine-tuning. OpenRAL is that infrastructure.

Typed runtime — eight well-defined layers connected by Pydantic v2 contracts. No magic globals, no hidden retries.
rSkill format — Hub repos containing weights, a rskill.yaml manifest, quantisation hints, latency budgets, and reproducible eval/. Install like a model.
Dual-system planning — a fast visuomotor policy (S1, 30–200 Hz) and a slow LLM planner (S2) emitting typed tool-calls. Replanning is bounded and explicit.
Safety kernel — deny-by-default, Python proposes, C++ disposes. ROSSafetyViolation is never silently caught.

Get started

# Install (no clone needed)
curl -fsSL https://raw.githubusercontent.com/AdrianLlopart/openral/master/scripts/install.sh | bash
openral doctor

# Browse and install a skill
openral rskill list
openral rskill install OpenRAL/rskill-smolvla-libero

# Run a simulated rollout
just sim-libero

# Deploy to hardware
openral deploy run --config deployments/so100_pickplace.yaml

Architecture

0 HAL Per-robot adapters — uniform connect / read_state / write_command Protocol shipped

1 Sensors SensorSpec → ROS 2 streams (RGB-D, F/T, IMU, lidar, tactile) shipped

2 World State tf2-aware 30 Hz snapshot; carries lifted detected_objects shipped

3 rSkill (S1) Fast visuomotor policy — VLA, 30–200 Hz, async action chunks shipped

4 Reasoning (S2) Slow LLM planner emitting typed ReasonerToolCall tool-calls partial

5 WAM World Action Model — mental simulation, failure anticipation planned

6 Safety C++ separate process, deny-by-default, certifiable, E-stop on fault partial

7 Observability OpenTelemetry spans + LeRobotDataset v3 flywheel shipped

What's in this organization

Type	What	Count
`rskill-*` models	Policy rSkills — VLA weights + manifest + eval	18
`rskill-*` models	Perception rSkills — RT-DETR object-detection	2
`dataset-*`	LeRobotDataset v3 demonstration datasets	growing

Policy rSkills

Each rSkill is a self-contained Hub repo: rskill.yaml manifest, model.safetensors weights, eval/ results, and a model card with runnable examples. Install with openral rskill install OpenRAL/rskill-<name>.

rSkill	Backbone	Target robot	License
`rskill-smolvla-libero`	SmolVLA	Franka Panda	Apache-2.0
`rskill-smolvla-metaworld`	SmolVLA	Rethink Sawyer	Apache-2.0
`rskill-smolvla-maniskill-franka`	SmolVLA × ManiSkill3	Franka Panda	Apache-2.0
`rskill-xvla-libero`	xVLA (Florence-2)	Franka Panda	Apache-2.0
`rskill-act-libero`	ACT	Franka Panda	Apache-2.0
`rskill-act-aloha`	ACT	ALOHA bimanual	MIT
`rskill-act-aloha-insertion`	ACT (peg insertion)	ALOHA bimanual	MIT
`rskill-diffusion-pusht`	Diffusion Policy	PushT 2-D	Apache-2.0
`rskill-molmoact2-libero-nf4`	MolmoAct2 NF4 (~5.5 B)	Franka Panda	Apache-2.0
`rskill-molmoact2-so101-nf4`	MolmoAct2 NF4	SO-101	Apache-2.0
`rskill-pi05-libero-nf4`	π0.5 NF4	Franka Panda	Research (weights non-Apache)
`rskill-pi05-so101-pickplace-nf4`	π0.5 NF4	SO-101	Research (weights)
`rskill-pi05-openarm-vision-nf4`	π0.5 NF4 bimanual	OpenArm v2	Apache-2.0
`rskill-pi05-robocasa365-human300-nf4`	π0.5 NF4	Panda Mobile	Research (weights)
`rskill-rldx1-pt-nf4`	RLDX-1 foundation (Qwen3-VL-8B, ~6.9 B)	Franka Panda	RLWRLD non-commercial
`rskill-rldx1-ft-libero-nf4`	RLDX-1 fine-tuned	Franka Panda	RLWRLD non-commercial
`rskill-rldx1-ft-gr1-nf4`	RLDX-1 (GR1 bimanual)	Fourier GR1	RLWRLD non-commercial
`rskill-rldx1-ft-simpler-widowx-nf4`	RLDX-1 (SimplerEnv)	WidowX	RLWRLD non-commercial

Perception rSkills

Emit ObjectsMetadata (2D detections lifted to 3D via depth) rather than an Action. Plug directly into the openral deploy graph.

rSkill	Backbone	Notes	License
`rskill-rtdetr-coco-r18`	RT-DETR R18	Lightweight ONNX	Apache-2.0
`rskill-rtdetr-v2-r50vd`	RT-DETR v2 R50vd	Higher accuracy	Apache-2.0

Free for any use Research-permissive weights (check upstream terms) Non-commercial — requires OPENRAL_ACCEPT_NONCOMMERCIAL=1

rSkill manifest format

# rskills/smolvla-libero/rskill.yaml (excerpt)
name: "OpenRAL/rskill-smolvla-libero"
version: "0.1.0"
license: "apache-2.0"
role: "s1"
embodiment_tags: ["franka_panda"]
sensors_required:
  - modality: "rgb"
    vla_feature_key: "observation.images.camera1"
latency_budget_ms: 120
quantization: null

Supported robots

Hardware + sim

SO-100 SO-101 ALOHA bimanual

Sim — HW bring-up in progress

Franka Panda UR5e UR10e

Sim (eval only)

Unitree H1 Unitree G1 Flexiv Rizon 4 Enactic OpenArm v2 Rethink Sawyer Fourier GR1 Panda Mobile Google Robot WidowX PushT 2D

Benchmarks

22 benchmark configs across LIBERO, MetaWorld, ManiSkill3, SimplerEnv, RoboCasa, gym-aloha, and gym-pusht. Every rSkill ships reproducible eval/<benchmark>.json you can regenerate locally:

openral benchmark run \
  --suite libero_spatial \
  --vla smolvla:rskill://OpenRAL/rskill-smolvla-libero

openral benchmark report

Observability & data flywheel

Every skill execution is an OpenTelemetry span — weights revision pinned, camera frames captured, LLM prompts logged. Traces replay as LeRobotDataset v3 rows, closing the loop from deployment back to training data.

openral dashboard   # OTLP receiver at :4318, live trace viewer

License

Everything in this organization is Apache-2.0 open core. A planned commercial tier (PolyForm SBL 1.0.0) will cover agentic orchestration features — free for teams under 100 people or <$1M revenue, and free for academic research at any size.

rSkill weights are governed by their upstream licenses (Apache-2.0 / MIT / research-permissive / RLWRLD non-commercial). The loader surfaces the posture at install time.