| 0 | PoC — Unitree R1 YOU ARE HERE
Validate the 3-layer architecture (Director → Managers → Execution) + the two loops (real-time correction and skill learning) on a real robot.
Deliverable: the control you already have — pattern proven on the R1. |
| 1 | Hardware-agnostic control layer (HAL)
Abstract the hardware behind a Hardware Abstraction Layer so Director/Managers/Execution run on any robot without rewriting the logic.
Deliverable: portable “control OS” + HAL. |
| 2 | OpenAI multi-agent brain
Implement the layers as OpenAI agents (planning, per-domain tools, real-time perception) and the skill memory as vector retrieval (RAG over skills).
Deliverable: plug-in “brain” service for the control OS. |
| 3 | Android integration
Choose the android vendor; port the HAL to its SDK; connect actuators and fuse sensors. Same brain, new body.
Deliverable: humanoid android driven by the same brain. |
| 4 | Embodied learning + safety
Skill acquisition by demonstration at scale; personal and shared memory; guardrails and a deterministic safety gate with certification.
Deliverable: validated skills + certified safety. |
| 5 | Product — the symbiosis
Android + OpenAI brain packaged as a single product: takes natural-language goals, acts in the physical world, and learns from the user.
Deliverable: commercial MVP. |