Reflex runs on rfx, our open-source AI-native robotics middleware. The
name of the SDK is
rfx — the platform that deploys and serves it is Reflex.What Reflex does
Reflex gives robots real-time access to datacenter-scale AI models. The stack is built around a single loop: simulate → collect → train → deploy → iterate Every step uses the same three-method robot contract —observe(), act(),
reset() — so a policy that works in simulation moves to hardware without
rewrites.
One primary CLI
rfx record, rfx deploy, rfx doctor. The public CLI surface is small
on purpose.Simulation as a primitive
Sim is first-class, not a plugin. Mock, Genesis, MuJoCo — same API.
Collection as a primitive
Recording structured observations into datasets is a core contract.
Self-describing models
Save once, load anywhere. HuggingFace Hub native.
Who it’s for
Researchers and engineers building AI policies for real robots. Reflex is not trying to recreate navigation, planning, SLAM, or a full autonomy stack. It is a compact core for modern AI robotics workloads where simulation, collection, policy execution, and real hardware share the same runtime contracts.How it’s built
- Rust core for runtime, transport, discovery, and hardware adapters
- Python SDK as the ergonomic surface for researchers
- Zenoh transport underneath — invisible plumbing, there when needed
- HuggingFace Hub native — push and pull policies like datasets
Quickstart
Record a demo and deploy a policy in under five minutes.
Architecture
The layers behind the loop.