A
Fragmentation, everywhere
Every robot SDK, every VLA, every sensor has its own dialect. OpenRAL puts a single, typed contract layer — Pydantic manifests over a ROS 2 backbone — between your robot and your models.
OPEN-SOURCE · APACHE-2.0 · BUILT ON ROS 2
The open harness
for physical AI.
OpenRAL is the typed orchestration layer that turns vision-language-action models, perception and reasoning into safe, runnable robot behavior — on real hardware and in simulation. One contract over many robots, many models, one safety boundary.
8typed layers
16robot manifests
30+ rSkills
63ADRs on record
RGB · head_cam
mug 0.94
bottle 0.88
DEPTH · stereo
bimanual · 16-DoF
S2 reasoner · trace
vlm
recall
navigate
pick
robometer
place
safety
01 — The problem
Robots, models and sensors don't speak the same language. OpenRAL is the contract that makes them.
B
Safety bolted on last
Safety is the architecture, not a wrapper. Python proposes actions; a deny-by-default supervisor disposes them. ROSSafetyViolation is never silently caught.
C
Models locked to vendors
Swap SmolVLA, π0.5, ACT, Diffusion Policy, MolmoAct2, GR00T N1.7 or RLDX-1 behind one Skill interface — embodiment tags, license posture and latency budgets enforced at load.
D
Nothing reproducible
Every execution is replayable from its trace alone: OpenTelemetry spans with tensor shapes, pinned weights, logged prompts, and a rosbag2 ↔ LeRobot dataset flywheel.
02 — What's inside
A full stack, composed — not reinvented.
OpenRAL builds on tf2, MoveIt 2, Nav2 and ros2_control, and adds the typed VLA / Skill / reasoning / safety / observability layer ROS 2 doesn't have.
03 — rSkills
Skills you can publish, score and share.
An rSkill is one robot capability, packaged as a HuggingFace Hub repo: an rskill.yaml manifest, the weights, an optional engine plan, and an eval/<benchmark>.json validated against a typed schema. Licenses are version-specific and enforced by the loader; latency budgets are contractual.
- Typed manifest — embodiment tags, capabilities, quantization, fallback skill
- License-governed — commercial posture derived from the weights' license
- Eval-scored — paper-cited or reproduced-locally, never faked
- Kinds —
vla·detector·vlm·reward·ros_action
name: "OpenRAL/rskill-pi05-openarm-vision"
kind: "vla"
role: "s1"
model_family: "pi05"
license: "apache-2.0"
embodiment_tags: ["openarm"]
actions: [PICK, PLACE, TRANSFER]
latency_budget:
warm_chunk_ms: 120
eval:
libero_spatial: 0.97204 — Who's building it
Built in the open.
Adrian Llopart
Founder · Senior AI Engineer, PhD
AI & humanoid robotics — vision-language-action models, perception, 3D human pose, and the harness that makes them run safely on real robots. Based in Copenhagen.
05 — Get in touch
Let's talk.
Building a robot, a fleet, or a model you want to run on OpenRAL? Researching the harness layer? Reach out — general questions go to hello@openral.com, partnerships to partner@openral.com.