Skip to main content
The design target is not “ROS2 but smaller” and not “a framework for one robot”. It is a narrow AI-native middleware built around the loop of simulate → collect → train → deploy → iterate.

​
Layers

The Rust core owns the middleware primitives:
  • Transport
  • Node / runtime lifecycle
  • Discovery
  • Hardware adapters
  • Typed wire contracts
  • Realtime and control utilities
The Python SDK is the ergonomic surface for researchers and robotics engineers:
  • Robot interface
  • Session
  • deploy
  • collection
  • Policy artifacts
  • Simulation helpers
Robot-specific integrations are adapters — they are not the identity of the framework.Examples: SO-101, Go2, G1, simulator backends, camera backends.
record, deploy, train, and runs are built on top of the core and SDK. They are useful entry points, but they are not the whole framework.

​
First-class primitives

Simulation

Sim is a primitive, not an extension. The same observation/action contract holds across mock, simulators, and real robots.

Collection

Collection is a primitive, not a sidecar script. Observations become structured datasets with stable metadata and artifact lineage.

Artifacts

Saved policies are self-describing and portable across machines and embodiments.

​
Non-goals

rfx is not trying to be:
  • a navigation stack
  • a SLAM framework
  • a planner library
  • a full autonomy platform
  • a robot-specific SDK disguised as middleware

​
Influence from Dora

The useful inspiration from Dora is process/dataflow discipline, clear runtime boundaries, low-latency transport, observability, and small, sharp middleware semantics. The goal is to bring that clarity into an AI-native robotics framework where simulation and collection are part of the base contract.