One capability at a time. Every subsystem earns its way onto the robot by passing a repeatable test behind a strict safety boundary. This is the same staged system taught in our books — here is the map.
Budget: ~$335–$550 · Gate: two weeks of reliable conversation
No moving parts. A Raspberry Pi 5 or Jetson Orin Nano Super, microphone array, speaker, camera, power, and one big physical stop button. You learn listening, speaking, memory, privacy, and recovery — almost everything that matters.
Exit test: it listens reliably, speaks clearly, and respects its mute button every single time.
Budget: ~$950–$2,150 · Gate: stable obstacle detection
Depth camera for obstacle detection, edge AI compute, an expressive display face, and pan/tilt gaze. Calibration, room mapping, and docking tests — plus the standing rule: no autonomous motion around children or pets during development.
Budget: ~$2,200–$4,300 · Gate: mapped rooms, reliable docking
A ROS 2 mobile base that maps your rooms and returns to its dock. Safety bumper, independent e-stop path, and a controlled test zone with mats and cones. A 30 kg robot is powered machinery, not a toy.
Budget: ~$1,700–$3,300 · Gate: low-force tasks under supervision
A single 6-axis learning arm with soft gripper, rigid mount, and an isolated safety relay. Fixtures and test objects make training repeatable. Force, torque, and speed stay bounded by the safety controller.
Budget: $25,000–$80,000+ all-in · Gate: vendor proves developer access in writing
Only now do you shop for a humanoid platform — and only after the vendor demonstrates the exact developer interface you need. Base models often lock you out of the access this project requires. Verify software access, warranty, and power compatibility before spending a single dollar. Freight, tax, spares, and training are part of the budget, not surprises.
The language model may suggest a goal — but a deterministic safety controller decides whether motion is allowed. The emergency stop, speed limits, and collision handling must keep working even if the network fails. Intelligence proposes. The safety controller disposes.
Low speed, cleared test zones, an adult spotter with direct access to the emergency stop. No exceptions during development.
Every stage ends with acceptance tests: cold boot, physical stop, voice stop, network loss, boundary respect, bumper response. No pass, no progress.