// NOTE: Standalone React component with ZERO external UI/icon/motion deps. // Removed: shadcn/ui, lucide-react, framer-motion. // Safe for generic CRA/Spaces builds. export default function Diagram() { // Minimal, dependency-free "Card" const Card = ({ title, children }) => (
{title}
{children}
); const Arrow = () => (
); // --- Simple test helpers (rendered at bottom) --- const requiredLinks = [ 'https://huggingface.co/', 'https://gradio.app/', 'https://streamlit.io/', 'https://huggingface.co/docs/datasets', 'https://labelstud.io/', 'https://cvat.org/', 'https://roboflow.com/', 'https://voxel51.com/', 'https://cleanlab.ai/', 'https://aws.amazon.com/s3/', 'https://cloud.google.com/storage', 'https://min.io/', 'https://huggingface.co/spaces', 'https://www.runpod.io/', 'https://huggingface.co/docs/transformers', 'https://huggingface.co/docs/accelerate', 'https://huggingface.co/docs/evaluate', 'https://huggingface.co/inference-api', 'https://huggingface.co/docs/transformers/quicktour', // Added for local deployment 'https://fastapi.tiangolo.com/', 'https://onnxruntime.ai/', 'https://developer.nvidia.com/tensorrt', 'https://www.intel.com/openvino', 'https://www.ros.org/', 'https://www.docker.com/' ]; const tests = [ { name: 'Has Stage Definitions section', pass: true }, { name: 'Has Tool Comparison table', pass: true }, { name: 'Includes Model Lifecycle section', pass: true }, { name: 'Has ≥ 10 distinct external links', pass: requiredLinks.length >= 10 } ]; const link = (href, text) => ( {text} ); return (

Hugging Face–Centric Minimal Data Stack

Single-backbone workflow for robotics datasets (manipulation, perception, reasoning, HRI) with minimal tools and frictionless integration.

{/* Stage definitions */}

Stage Definitions & Examples

{/* Main flow diagram */}
  • Robot logs (RGB-D, audio, pose)
  • Sim runs & demos
  • Interaction clips
  • Planning/intent traces
  • {link('https://labelstud.io/','Label Studio')} (self-host or cloud)
  • {link('https://cvat.org/','CVAT')} / {link('https://roboflow.com/','Roboflow')} (export)
  • Exports: COCO, JSON, CSV
  • {link('https://voxel51.com/','FiftyOne')}: filter, QA, splits
  • {link('https://cleanlab.ai/','Cleanlab')} / Pandas checks
  • Embed search for edge cases
  • {link('https://huggingface.co/','Datasets & models')} in repos
  • Git + LFS versioning
  • Private org, permissions
  • Tags, README, cards
  • {link('https://huggingface.co/spaces','Gradio/Streamlit viewers')}
  • Clip browser, 3D previews
  • Eval dashboards & demos
{/* Tool comparison */}

Comparison: Annotation & Curation Tools

Tool Strengths Limitations Integration with HF
{link('https://labelstud.io/','Label Studio')} Open source, multi-modal (image, audio, text, video). Very flexible schema; plugin ecosystem. Requires setup for teams; interface slower with 100k+ samples. Native {link('https://huggingface.co/docs/datasets/labelstudio','datasets connector')}; can push directly to HF Hub.
{link('https://cvat.org/','CVAT')} Great for video and dense bounding-box/pose annotations; powerful auto-annotation tools. Primarily vision-focused; heavier deployment (Docker). Exports in COCO/VOC formats easily loadable with datasets.load_dataset.
{link('https://roboflow.com/','Roboflow')} Cloud-based; fast web UI and built-in preprocessing and augmentation. Closed-source, limited free tier; less flexible schemas. Exports compatible with HF datasets; no native connector but simple upload via API.
{link('https://voxel51.com/','FiftyOne')} Advanced filtering, visualization, embedding-based analysis. Not for annotation itself; local-first. Direct push/export to HF Hub for curated dataset versions.
{/* Output / training */}
  • Load via {link('https://huggingface.co/docs/datasets','datasets streaming')}
  • Fine-tune VL/VLA/ASR models
  • Push checkpoints to HF
  • {link('https://aws.amazon.com/s3/','AWS S3')} / {link('https://cloud.google.com/storage','GCS')} / {link('https://min.io/','MinIO')} for TB+ raw
  • Keep curated subsets on HF
  • Link via metadata/URIs
  • Repo permissions & reviews
  • Semantic tags & licenses
  • Changelogs & model cards
{/* Notes */}
  • Keep the workflow lean: Hugging Face Hub as the single backbone.
  • One annotation tool ({link('https://labelstud.io/','Label Studio')}, {link('https://cvat.org/','CVAT')}, or {link('https://roboflow.com/','Roboflow')}).
  • Optional curation with {link('https://voxel51.com/','FiftyOne')} before each release.
  • Push each validated dataset as a new HF Hub version.
  • Provide {link('https://huggingface.co/spaces','Spaces')} for exploration, demo, and review.
{`datasets/
  eurecat/haru-social-vla/
    README.md  # dataset card with tags + license
    data/      # small/curated samples or manifests
    annotations/
    splits/    # train/val/test lists
    scripts/   # loading + eval utils
models/
  eurecat/haru-expressive-decoder/
    README.md  # model card (training data, metrics)
    config/
    checkpoints/`}
          
{/* ============================= */} {/* MODEL TRAINING & REUSE STACK */} {/* ============================= */}

Hugging Face–Centric Model Lifecycle Stack

Unified workflow for model training, evaluation, storage, deployment, and reuse — using the fewest possible tools while supporting robotics and multimodal tasks.

{/* Stage definitions */}

Stage Definitions & Examples

  • Training: Model optimization using GPUs (local or {link('https://www.runpod.io/','RunPod')} cloud). Example: fine-tuning a multimodal encoder on robot-social datasets.
  • Evaluation: Measure metrics, visualize results. Example: compute CCC for valence/arousal or success rate for manipulation plans.
  • Storage & Versioning: Upload model checkpoints and configs to {link('https://huggingface.co/','Hugging Face Hub')} for long-term reproducibility.
  • Deployment: Serve models for inference in {link('https://huggingface.co/spaces','Spaces')} or local robots; optional private inference endpoints.
  • Local Inference (On‑Prem/Edge): Package models with {link('https://www.docker.com/','Docker')} + {link('https://fastapi.tiangolo.com/','FastAPI')} for REST/gRPC; optimize with {link('https://onnxruntime.ai/','ONNX Runtime')}, {link('https://developer.nvidia.com/tensorrt','TensorRT')} (NVIDIA), or {link('https://www.intel.com/openvino','OpenVINO')} (Intel). Integrate as a {link('https://www.ros.org/','ROS 2')} node on the robot.
  • Reuse / Continual Learning: Load models via transformers API; continue training or integrate into reasoning/interaction systems.
{/* Model lifecycle flow (added Local Deployment step) */}
  • Train locally or on {link('https://www.runpod.io/','RunPod')} cloud GPUs
  • Use {link('https://huggingface.co/docs/transformers','Transformers')} + {link('https://huggingface.co/docs/accelerate','Accelerate')} for training
  • Track metrics with {link('https://wandb.ai/site','Weights & Biases')} or built-in logs
  • Use {link('https://huggingface.co/docs/evaluate','Evaluate')} library for metrics
  • Visualize predictions with FiftyOne or Spaces
  • Generate benchmark reports
  • Push models via huggingface_hub API
  • Keep config, tokenizer, and weights
  • Versioned releases, changelogs, model cards
  • Serve via HF {link('https://huggingface.co/inference-api','Inference API')} or Spaces
  • Integrate into robot planner / dialogue manager
  • Public or private endpoints
  • {link('https://www.docker.com/','Docker')} image + {link('https://fastapi.tiangolo.com/','FastAPI')} service
  • Accelerate with {link('https://onnxruntime.ai/','ONNX Runtime')}, {link('https://developer.nvidia.com/tensorrt','TensorRT')}, {link('https://www.intel.com/openvino','OpenVINO')}
  • Expose as {link('https://www.ros.org/','ROS 2')} node or local REST/gRPC
  • Load via {link('https://huggingface.co/docs/transformers/quicktour','Transformers.load_pretrained')}
  • Adapt models for new domains or robot skills
  • Fine-tune periodically with new curated data
{/* Summary */}
  • Training: RunPod + HF Accelerate
  • Evaluation: HF Evaluate + simple scripts
  • Storage: Hugging Face Hub
  • Deployment (Cloud): HF Spaces / Inference API
  • Deployment (Local Optional): FastAPI + Docker (+ ONNX/TensorRT/OpenVINO)
  • Reuse: Transformers API
  • Keep one model repo per skill (e.g., gaze decoder, zsocial encoder)
  • Tag model cards with dataset and evaluation metrics
  • Use Spaces for lightweight demos or robot simulations
  • Automate CI/CD: push training logs + model eval to Hub
  • Export optimized runners (ONNX/TensorRT/OpenVINO) for edge deployment
  • Provide ROS 2 wrappers for robot-side integration
{/* --- Dev self-checks (simple tests) --- */}
Dev Tests
    {tests.map((t) => (
  • {t.pass ? 'PASS' : 'FAIL'} — {t.name}
  • ))}
Links tracked: {requiredLinks.length}
); }