Get started

# Install (no clone needed)
curl -fsSL https://raw.githubusercontent.com/OpenRAL/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

What's in this organization

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>. The table below shows a representative subset of the 20 VLA policy rSkills; classical motion planners ship as separate ros_action rSkills (see below).

ROS-action rSkills

kind: ros_action. Classical motion planners packaged as rSkills, so the S2 reasoner dispatches a MoveIt or Nav2 motion exactly like a learned policy (this is the "VLA and classical controllers" thesis). No weights; they wrap a ROS 2 action server.

Perception detector rSkills

kind: detector. Emit ObjectsMetadata (2D detections lifted to 3D via depth) rather than an Action. Plug directly into the openral deploy graph and fold into World State.

Scene-understanding VLM rSkill

kind: vlm. A scene VLM that answers open-ended questions about the current view. It powers the S2 reasoner's read-only query_scene tool for task-progress / success verification ("did the grasp succeed?"); it holds no actuation authority (ADR-0047). Runs out-of-process in an NF4 sidecar served by scene_vlm_node.

Reward-monitor rSkill

kind: reward. A robotic reward / progress-monitor model that runs in parallel with a VLA and scores the live rollout, emitting per-frame normalised progress (0–1) and success probability. It powers the S2 reasoner's read-only query_task_progress tool ("is the task succeeding right now?"); advisory, never on the control path (ADR-0057). Runs out-of-process in an NF4 sidecar served by reward_monitor_node; the pre-quantized checkpoint is meta-loaded directly as 4-bit.

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

Benchmarks

12 benchmark scenes 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. The entire codebase ships under the same permissive license.

rSkill weights are governed by their own upstream licenses (Apache-2.0 / MIT / research-permissive / RLWRLD non-commercial). The loader surfaces each weight's posture at install time. This is third-party license lineage for models OpenRAL does not own; it does not affect the Apache-2.0 license of OpenRAL's own code.