Quickstart

Smiling Buddha is a set of open tools, datasets, and benchmarks for physical AI research. This guide gets you from zero to running a benchmark in under ten minutes.

Install

Install the CLI and library from PyPI:

pip install smiling-buddha

Or from source, if you want the latest working branch:

git clone https://github.com/smilingbuddha/smilingbuddha
cd smilingbuddha
pip install -e .

First run

Grab a small sample dataset, run the reference baseline, and get your first evaluation score:

sb datasets pull SB-Perception-01-sample
sb benchmark run SB-Perception-01 --baseline reference
sb benchmark score

Expected output looks something like:

SB-Perception-01 · reference baseline
─────────────────────────────────────
Precision  0.842
Recall     0.791
F1         0.816
V-score    V2 (validated on 3 of 4 splits)

Data model

Every dataset in the Smiling Buddha library follows the LeRobot v2.1 native format. This gives you three things for free:

• Immediate compatibility with the lerobot Python library
• Consistent episode / observation / action structure across all datasets
• Version-safe schema evolution — old datasets keep working with new library versions

V0 → V3 validation taxonomy

Every artifact we publish carries a validation score from V0 (concept-only, unvalidated) to V3 (validated on real hardware, at deployment fidelity). Full definitions:

• V0 — Conceptual. Documented approach, not yet implemented.
• V1 — Simulation-only. Implemented and evaluated in pure simulation.
• V2 — Real-data validated. Evaluated on captured real-world sequences.
• V3 — Deployment-validated. Evaluated on live deployment traces at production fidelity.

API reference

The Python API is small on purpose. Three entry points cover 90 % of use cases: sb.datasets, sb.benchmarks, sb.tools. Full reference in the source repository.