← All projects

Personal project

Robotic Reasoning Models for Physical AI

RRM-1 is a private personal project for robot-agnostic embodied reasoning, connecting persistent world modeling, structured task planning, independent supervision, and embodiment-aware execution.

RRM-1 is designed to make robot behavior inspectable across the full task loop. Rather than sending a high-level goal directly to a policy, it maintains structured state about objects, relations, actions, and expected effects, then checks proposed actions before execution.

Current work emphasizes explicit contracts and replayable evidence: versioned state, bounded planning calls, admission checks, safety gates, and monitored execution traces. The architecture is designed to evaluate pluggable models, reasoning modules, and embodiment adapters through stable interfaces rather than binding the system to one model or robot.

High-level operating model

01

Perception & state evidence

Observations · Confidence · Object and relation evidence

02

Persistent world model

Objects · Relations · Task context · Expected effects

03

Task reasoning & replanning

Structured task graphs · Action proposals · Recovery decisions

04

Independent supervision

Admission · Approval · Safety and feasibility checks

05

Embodiment-aware execution

Adapter mapping · Monitored action · Observation and review

Only admitted actions cross into execution; new evidence informs subsequent verification, recovery, or replanning.

This is an intentionally high-level public overview. The implementation, experimental setup, evaluation details, and current results remain private while the research develops.