| <!DOCTYPE html> |
| <html lang="en"> |
| <head> |
| <meta charset="UTF-8" /> |
| <meta name="viewport" content="width=device-width, initial-scale=1.0" /> |
| <title>OpenRAL: Runtime for VLA Robot Agents</title> |
| <style> |
| *, *::before, *::after { box-sizing: border-box; margin: 0; padding: 0; } |
| |
| :root { |
| --black: #0f0f0f; |
| --off-white: #f8f7f4; |
| --mid: #6b6b6b; |
| --border: #e0ddd8; |
| --accent: #1a1a1a; |
| --code-bg: #f1f0ed; |
| --green: #16a34a; |
| --yellow: #ca8a04; |
| --red: #dc2626; |
| } |
| |
| body { |
| font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Helvetica, Arial, sans-serif; |
| background: var(--off-white); |
| color: var(--black); |
| line-height: 1.6; |
| font-size: 15px; |
| } |
| |
| a { color: var(--black); text-underline-offset: 3px; } |
| a:hover { opacity: 0.65; } |
| |
| |
| .page { max-width: 860px; margin: 0 auto; padding: 48px 24px 80px; } |
| |
| |
| .header { |
| display: flex; |
| align-items: center; |
| gap: 20px; |
| margin-bottom: 32px; |
| padding-bottom: 32px; |
| border-bottom: 1px solid var(--border); |
| } |
| .header img { width: 56px; height: auto; flex-shrink: 0; } |
| .header-text h1 { font-size: 28px; font-weight: 700; letter-spacing: -0.5px; } |
| .header-text p { color: var(--mid); margin-top: 4px; font-size: 15px; max-width: 580px; } |
| |
| |
| .badges { display: flex; flex-wrap: wrap; gap: 8px; margin: 20px 0 32px; } |
| .badge { |
| display: inline-flex; |
| align-items: center; |
| gap: 5px; |
| padding: 4px 10px; |
| border: 1px solid var(--border); |
| border-radius: 20px; |
| font-size: 12px; |
| font-weight: 500; |
| text-decoration: none; |
| color: var(--black); |
| background: white; |
| white-space: nowrap; |
| } |
| .badge:hover { background: var(--code-bg); opacity: 1; } |
| |
| |
| h2 { font-size: 18px; font-weight: 700; margin: 40px 0 16px; letter-spacing: -0.3px; } |
| h3 { font-size: 14px; font-weight: 600; text-transform: uppercase; letter-spacing: 0.5px; color: var(--mid); margin: 28px 0 12px; } |
| p { margin-bottom: 12px; color: #2a2a2a; } |
| |
| |
| .pitch { |
| background: white; |
| border: 1px solid var(--border); |
| border-radius: 10px; |
| padding: 24px; |
| margin-bottom: 32px; |
| } |
| .pitch ul { padding-left: 20px; } |
| .pitch li { margin-bottom: 8px; } |
| .pitch li strong { font-weight: 600; } |
| |
| |
| pre { |
| background: var(--code-bg); |
| border: 1px solid var(--border); |
| border-radius: 8px; |
| padding: 16px 20px; |
| font-family: "SF Mono", "Fira Code", "Cascadia Code", monospace; |
| font-size: 13px; |
| line-height: 1.65; |
| overflow-x: auto; |
| margin-bottom: 24px; |
| } |
| code { font-family: inherit; } |
| p code, li code, td code { |
| background: var(--code-bg); |
| border: 1px solid var(--border); |
| border-radius: 4px; |
| padding: 1px 5px; |
| font-size: 12.5px; |
| } |
| |
| |
| .table-wrap { overflow-x: auto; margin-bottom: 24px; } |
| table { width: 100%; border-collapse: collapse; font-size: 13.5px; } |
| th { |
| text-align: left; |
| padding: 8px 12px; |
| border-bottom: 2px solid var(--border); |
| font-weight: 600; |
| color: var(--mid); |
| font-size: 12px; |
| text-transform: uppercase; |
| letter-spacing: 0.4px; |
| } |
| td { padding: 9px 12px; border-bottom: 1px solid var(--border); vertical-align: top; } |
| tr:last-child td { border-bottom: none; } |
| tr:hover td { background: #faf9f6; } |
| |
| |
| .lic { display: inline-block; width: 8px; height: 8px; border-radius: 50%; margin-right: 4px; vertical-align: middle; } |
| .lic-green { background: var(--green); } |
| .lic-yellow { background: var(--yellow); } |
| .lic-red { background: var(--red); } |
| |
| |
| .robot-group { margin-bottom: 20px; } |
| .robot-group h3 { margin-top: 0; } |
| .robot-chips { display: flex; flex-wrap: wrap; gap: 6px; } |
| .chip { |
| padding: 4px 10px; |
| border-radius: 16px; |
| font-size: 12.5px; |
| font-weight: 500; |
| border: 1px solid var(--border); |
| background: white; |
| } |
| .chip-green { border-color: #86efac; background: #f0fdf4; } |
| .chip-blue { border-color: #93c5fd; background: #eff6ff; } |
| .chip-gray { background: var(--code-bg); } |
| |
| |
| .footer { |
| margin-top: 56px; |
| padding-top: 24px; |
| border-top: 1px solid var(--border); |
| display: flex; |
| flex-wrap: wrap; |
| gap: 20px; |
| font-size: 13px; |
| color: var(--mid); |
| } |
| .footer a { color: var(--mid); } |
| |
| @media (max-width: 600px) { |
| .header { flex-direction: column; align-items: flex-start; gap: 12px; } |
| } |
| </style> |
| </head> |
| <body> |
| <div class="page"> |
|
|
| |
| <div class="header"> |
| <img src="/spaces/OpenRAL/README/resolve/main/openral-logo.svg" alt="OpenRAL logo" /> |
| <div class="header-text"> |
| <h1>OpenRAL</h1> |
| <p>An open-source operating layer for embodied AI, OpenRAL unifies fast policies, slow reasoning, and classical control into one typed, traceable, safety-first runtime for deployable robot agents.</p> |
| </div> |
| </div> |
|
|
| |
| <div class="badges"> |
| <a class="badge" href="https://github.com/OpenRAL/openral">⬛ GitHub</a> |
| <a class="badge" href="https://openral.github.io/openral/">📄 Docs</a> |
| <a class="badge" href="https://discord.gg/3paXT2bVyB">💬 Discord</a> |
| <span class="badge">Apache-2.0</span> |
| <span class="badge">ROS 2 Jazzy</span> |
| <span class="badge">Python 3.12</span> |
| <span class="badge">lerobot 0.6.0</span> |
| </div> |
|
|
| |
| <div class="pitch"> |
| <p>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. <strong>OpenRAL is that infrastructure.</strong></p> |
| <br/> |
| <ul> |
| <li><strong>Typed runtime</strong>: eight well-defined layers connected by Pydantic v2 contracts. No magic globals, no hidden retries.</li> |
| <li><strong>rSkill format</strong>: Hub repos containing weights, a <code>rskill.yaml</code> manifest, quantisation hints, latency budgets, and reproducible <code>eval/</code>. Install like a model.</li> |
| <li><strong>Dual-system planning</strong>: a fast visuomotor policy (S1, 30–200 Hz) and a slow LLM planner (S2) emitting typed tool-calls. Replanning is bounded and explicit.</li> |
| <li><strong>Safety kernel</strong>: deny-by-default, Python proposes, C++ disposes. <code>ROSSafetyViolation</code> is never silently caught.</li> |
| </ul> |
| </div> |
|
|
| |
| <h2>Get started</h2> |
| <pre><code># 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</code></pre> |
|
|
| |
| <h2>Policy rSkills</h2> |
| <p>Each rSkill is a self-contained Hub repo: <code>rskill.yaml</code> manifest, <code>model.safetensors</code> weights, <code>eval/</code> results, and a model card with runnable examples. Install with <code>openral rskill install OpenRAL/rskill-<name></code>. The table below shows a representative subset of the 19 VLA policy rSkills; classical motion planners ship as separate <code>ros_action</code> rSkills (see below).</p> |
|
|
| <div class="table-wrap"> |
| <table> |
| <thead><tr><th>rSkill</th><th>Backbone</th><th>Target robot</th><th>License</th></tr></thead> |
| <tbody> |
| <tr> |
| <td><code>rskill-smolvla-libero</code></td> |
| <td>SmolVLA</td><td>Franka Panda</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-smolvla-metaworld</code></td> |
| <td>SmolVLA</td><td>Rethink Sawyer</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-smolvla-maniskill-franka</code></td> |
| <td>SmolVLA × ManiSkill3</td><td>Franka Panda</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-xvla-libero</code></td> |
| <td>xVLA (Florence-2)</td><td>Franka Panda</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-act-libero</code></td> |
| <td>ACT</td><td>Franka Panda</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-act-aloha</code></td> |
| <td>ACT</td><td>ALOHA bimanual</td> |
| <td><span class="lic lic-green"></span>MIT</td> |
| </tr> |
| <tr> |
| <td><code>rskill-act-aloha-insertion</code></td> |
| <td>ACT (peg insertion)</td><td>ALOHA bimanual</td> |
| <td><span class="lic lic-green"></span>MIT</td> |
| </tr> |
| <tr> |
| <td><code>rskill-diffusion-pusht</code></td> |
| <td>Diffusion Policy</td><td>PushT 2-D</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-molmoact2-libero-nf4</code></td> |
| <td>MolmoAct2 NF4 (~5.5 B)</td><td>Franka Panda</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-molmoact2-so101-nf4</code></td> |
| <td>MolmoAct2 NF4</td><td>SO-101</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-smolvla-robotwin</code></td> |
| <td>SmolVLA × RoboTwin 2.0</td><td>ALOHA AgileX (dual-arm)</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-smolvla-so101-pick-place-pen</code></td> |
| <td>SmolVLA</td><td>SO-101</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-3d-diffuser-actor-rlbench</code></td> |
| <td>3D Diffuser Actor</td><td>Franka Panda (RLBench)</td> |
| <td><span class="lic lic-green"></span>MIT</td> |
| </tr> |
| <tr> |
| <td><code>rskill-rldx1-ft-rc365-nf4</code></td> |
| <td>RLDX-1 (RoboCasa 365)</td><td>Panda Mobile</td> |
| <td><span class="lic lic-red"></span>RLWRLD non-commercial</td> |
| </tr> |
| <tr> |
| <td><code>rskill-rldx1-ft-libero-nf4</code></td> |
| <td>RLDX-1 fine-tuned</td><td>Franka Panda</td> |
| <td><span class="lic lic-red"></span>RLWRLD non-commercial</td> |
| </tr> |
| <tr> |
| <td><code>rskill-rldx1-ft-gr1-nf4</code></td> |
| <td>RLDX-1 (GR1 bimanual)</td><td>Fourier GR1</td> |
| <td><span class="lic lic-red"></span>RLWRLD non-commercial</td> |
| </tr> |
| <tr> |
| <td><code>rskill-rldx1-ft-simpler-widowx-nf4</code></td> |
| <td>RLDX-1 (SimplerEnv)</td><td>WidowX</td> |
| <td><span class="lic lic-red"></span>RLWRLD non-commercial</td> |
| </tr> |
| <tr> |
| <td><code>rskill-gr00t-n17-libero</code></td> |
| <td>GR00T N1.7</td><td>Franka Panda</td> |
| <td><span class="lic lic-green"></span>NVIDIA Open Model</td> |
| </tr> |
| <tr> |
| <td><code>rskill-openvla-oft-simpler-widowx-nf4</code></td> |
| <td>OpenVLA-OFT NF4 (RLinf PPO, ManiSkill3)</td><td>WidowX</td> |
| <td><span class="lic lic-green"></span>MIT</td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
|
|
| <h2>ROS-action rSkills</h2> |
| <p><code>kind: ros_action</code>. 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 <em>and</em> classical controllers" thesis). No weights; they wrap a ROS 2 action server.</p> |
| <div class="table-wrap"> |
| <table> |
| <thead><tr><th>rSkill</th><th>Planner</th><th>What it does</th><th>License</th></tr></thead> |
| <tbody> |
| <tr> |
| <td><code>rskill-moveit-eef-pose</code></td> |
| <td>MoveIt MoveGroup</td><td>Collision-free motion to a 6-DOF Cartesian end-effector pose (e.g. a pre-grasp)</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-moveit-joints</code></td> |
| <td>MoveIt MoveGroup</td><td>Collision-free motion to a target joint configuration (e.g. a policy's in-distribution start pose)</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-moveit-look-at</code></td> |
| <td>MoveIt MoveGroup</td><td>Aim a wrist-mounted camera at a 3-D point so a later perception query / grasp sees the object framed</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-nav2-navigate-to-pose</code></td> |
| <td>Nav2 NavigateToPose</td><td>Drive a mobile base to an absolute (<code>map</code>) or relative (<code>base_link</code>) goal pose</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
|
|
| <h2>Perception detector rSkills</h2> |
| <p><code>kind: detector</code>. Emit <code>ObjectsMetadata</code> (2D detections lifted to 3D via depth) rather than an <code>Action</code>. Plug directly into the <code>openral deploy</code> graph and fold into World State.</p> |
| <div class="table-wrap"> |
| <table> |
| <thead><tr><th>rSkill</th><th>Backbone</th><th>Notes</th><th>License</th></tr></thead> |
| <tbody> |
| <tr> |
| <td><code>rskill-rtdetr-coco-r18</code></td> |
| <td>RT-DETR R18</td><td>Lightweight ONNX</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-rtdetr-v2-r50vd</code></td> |
| <td>RT-DETR v2 R50vd</td><td>Higher accuracy</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-omdet-turbo-indoor</code></td> |
| <td>OmDet-Turbo (Swin-tiny)</td><td>Real-time open-vocab; ~230-class indoor vocabulary, unprompted background producer</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-omdet-turbo-locator</code></td> |
| <td>OmDet-Turbo (Swin-tiny)</td><td>On-demand <code>locate_in_view</code>: lightweight in-process "find X"</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-locateanything-3b-nf4</code></td> |
| <td>LocateAnything-3B</td><td>Open vocabulary (out-of-process NF4 sidecar)</td> |
| <td><span class="lic lic-red"></span>NVIDIA non-commercial</td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
|
|
| <h2>Scene-understanding VLM rSkill</h2> |
| <p><code>kind: vlm</code>. A scene VLM that answers open-ended questions about the current view. It powers the S2 reasoner's <strong>read-only</strong> <code>query_scene</code> 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 <code>scene_vlm_node</code>.</p> |
| <div class="table-wrap"> |
| <table> |
| <thead><tr><th>rSkill</th><th>Backbone</th><th>Notes</th><th>License</th></tr></thead> |
| <tbody> |
| <tr> |
| <td><code>rskill-qwen35-4b-nf4</code></td> |
| <td>Qwen3.5-4B NF4 (~2.5 GB)</td><td>Reasoner <code>query_scene</code> tool · role <code>s2</code></td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
|
|
| <h2>Reward-monitor rSkills</h2> |
| <p><code>kind: reward</code>. A robotic <strong>reward / progress-monitor</strong> model that runs <strong>in parallel with a VLA</strong> and scores the live rollout, emitting per-frame normalised progress (0–1) and success probability. It powers the S2 reasoner's <strong>read-only</strong> <code>query_task_progress</code> 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 <code>reward_monitor_node</code>; the pre-quantized checkpoint is meta-loaded directly as 4-bit.</p> |
| <div class="table-wrap"> |
| <table> |
| <thead><tr><th>rSkill</th><th>Backbone</th><th>Notes</th><th>License</th></tr></thead> |
| <tbody> |
| <tr> |
| <td><code>rskill-robometer-4b-nf4</code></td> |
| <td>Robometer-4B / Qwen3-VL-4B NF4 (~3.3 GB)</td><td>Reasoner <code>query_task_progress</code> tool · role <code>s2</code> · fits 8 GB alongside a small VLA</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| <tr> |
| <td><code>rskill-topreward-qwen3vl-4b-nf4</code></td> |
| <td>TOPReward (zero-shot) / Qwen3-VL-4B NF4 (3.13 GB)</td><td>Reasoner <code>query_task_progress</code> tool · role <code>s2</code> · per-frame progress via lerobot prefix sweep · fits 8 GB</td> |
| <td><span class="lic lic-green"></span>Apache-2.0</td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
|
|
| |
| <p style="font-size:12.5px; color: var(--mid);"> |
| <span class="lic lic-green"></span> Free for any use |
| <span class="lic lic-yellow"></span> Research-permissive weights (check upstream terms) |
| <span class="lic lic-red"></span> Non-commercial: requires <code>OPENRAL_ACCEPT_NONCOMMERCIAL=1</code> |
| </p> |
|
|
| |
| <h2>rSkill manifest format</h2> |
| <pre><code># 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</code></pre> |
|
|
| |
| <h2>Supported robots</h2> |
|
|
| <div class="robot-group"> |
| <h3>Hardware + sim</h3> |
| <div class="robot-chips"> |
| <span class="chip chip-green">SO-100</span> |
| <span class="chip chip-green">SO-101</span> |
| <span class="chip chip-green">ALOHA bimanual</span> |
| </div> |
| </div> |
|
|
| <div class="robot-group"> |
| <h3>Sim: HW bring-up in progress</h3> |
| <div class="robot-chips"> |
| <span class="chip chip-blue">Franka Panda</span> |
| <span class="chip chip-blue">UR5e</span> |
| <span class="chip chip-blue">UR10e</span> |
| </div> |
| </div> |
|
|
| <div class="robot-group"> |
| <h3>Sim (eval only)</h3> |
| <div class="robot-chips"> |
| <span class="chip chip-gray">Unitree H1</span> |
| <span class="chip chip-gray">Unitree G1</span> |
| <span class="chip chip-gray">Flexiv Rizon 4</span> |
| <span class="chip chip-gray">Enactic OpenArm v2</span> |
| <span class="chip chip-gray">Anvil OpenARM 2.0</span> |
| <span class="chip chip-gray">Rethink Sawyer</span> |
| <span class="chip chip-gray">Fourier GR1</span> |
| <span class="chip chip-gray">ALOHA AgileX (dual-arm)</span> |
| <span class="chip chip-gray">Panda Mobile</span> |
| <span class="chip chip-gray">Google Robot</span> |
| <span class="chip chip-gray">WidowX</span> |
| <span class="chip chip-gray">PushT 2D</span> |
| </div> |
| </div> |
|
|
| |
| <h2>Benchmarks</h2> |
| <p>12 benchmark scenes across LIBERO, MetaWorld, ManiSkill3, SimplerEnv, RoboCasa, gym-aloha, and gym-pusht. Every rSkill ships reproducible <code>eval/<benchmark>.json</code> you can regenerate locally:</p> |
| <pre><code>openral benchmark run \ |
| --suite libero_spatial \ |
| --vla smolvla:rskill://OpenRAL/rskill-smolvla-libero |
|
|
| openral benchmark report</code></pre> |
|
|
| |
| <h2>Observability & data flywheel</h2> |
| <p>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.</p> |
| <pre><code>openral dashboard # OTLP receiver at :4318, live trace viewer</code></pre> |
|
|
| |
| <h2>License</h2> |
| <p>Everything in this organization is <strong>Apache-2.0</strong>. The entire codebase ships under the same permissive license.</p> |
| <p>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.</p> |
|
|
| |
| <div class="footer"> |
| <a href="https://github.com/OpenRAL/openral">GitHub</a> |
| <a href="https://openral.github.io/openral/">Docs</a> |
| <a href="https://discord.gg/3paXT2bVyB">Discord</a> |
| <a href="mailto:hello@openral.com">Email</a> |
| <span>Apache-2.0 · © OpenRAL contributors</span> |
| </div> |
|
|
| </div> |
| </body> |
| </html> |
