The Work

Seventy-four stages.
One invariant.

NEURIX is built in stages — each one an architectural decision, validated in simulation and, where applicable, on a real humanoid. Below is the shape of the work so far.

By the Numbers

Validated Capabilities

Capabilities are considered validated when they pass under the NEURIX runtime — unit tests, live CLI path, and an operator scenario. Otherwise they are partial.

• Stage 66 · v1.39
  First MuJoCo-native walking policy.

  6.5 m in 15 s. Zero falls. Trained directly on the canonical 43DOF body with correct actuator gains — the first policy to survive the transition from training physics to deployment physics.

• Stage 73 · v1.46
  Law #0 enforced end-to-end.

  Autonomous teleport-standup removed. The robot can no longer be silently reset to an upright pose. Recovery now requires a real get-up policy — or an operator override that logs the violation.

• Stage 74
  Imitation-bootstrap pipeline.

  Demonstrations → behaviour cloning → PPO fine-tune, validated end-to-end. BC+PPO outperforms from-scratch in regime-aligned setups. Four architectural laws (T-1…T-4) formalised from failures along the way.

• Stage 71 · v1.44
  Law #0 perception.

  The robot sees only through sensors — IMU odometry, LIDAR, camera through a vision-language model. No qpos reads in runtime code. Simulator and reality read the same channels.

• Stage 70 · v1.43
  CP-resolved profiles.

  Every motion resolves its embodiment profile through the Control Plane. No bypass paths. Full audit closed — 100% of commands travel the canonical pipeline.

• Stage 68–69
  Embodiment-agnostic command layer.

  Motion commands are specified against an abstract body. Profile adapters translate them to a specific robot. The same skill runs against different adapters without modification.

• Stage 49 · v1.36
  Canonical 43DOF embodiment, frozen.

  Unitree G1 with Dex3-1 hands, 43 degrees of freedom. Kinematics, actuator gains, and observation/action contracts are frozen as the reference body for all downstream training.

• Stage 42
  Navigation on verified locomotion.

  Goal-directed navigation runs on top of a locomotion policy that has been validated independently. Path planning and body control are cleanly separated.

• Stage 24–26
  Kinematic planning with Dubins paths.

  Planner respects body kinematic limits. Obstacle-aware A* on an envelope-constrained grid. 100% repeatable on clear-path routes; obstacle routes improved from 0% to ~40% reliability.

• Stage 17 · v1.7
  Embodiment adapter architecture.

  Skills describe intent against a capability surface. Adapters bind that surface to specific robots. This is the architecture that makes every stage above possible.

• Stage 9–10
  Closed perception loop.

  sense → provider → facts → evaluation → plan → propose → approve → execute. The canonical cycle of the runtime.

Stage Evolution

Work groups into architectural eras. Each era resolves one class of problem and makes the next one addressable.

Active Runtime

The current baseline deployed under the runtime.

What Holds It Together

Two laws of the project and six laws of humanoid synthesis, learned on the way.

Continue