A 5'9", 29-joint bipedal humanoid robot, designed and built from scratch: no kit, no off-the-shelf frame. It's an anticipatory personal assistant here on Earth, and a step toward robotics for space.
Michael J. Caruana, independent robotics engineer mjcaruana.com · GitHub · Support the build
| Scale | 5'9" (1.75 m), 29 actuated joints including articulated hands |
| Structure | 3D-printed carbon-fiber composite, designed in CAD |
| Electronics | Custom-designed circuit boards; distributed real-time controllers on a high-speed CAN-FD bus |
| Intelligence | LLM-based reasoning layer over deterministic real-time control; on-device learning |
| Method | Every architecture decision is backed by a measured, published experiment |
Latest result: Exp01, IMU Bridge Latency
Measured the link between the robot's real-time controller and its Linux "thinking" processor across three runs and 153,808 samples:
- About 6 ms typical, up to about 23 ms worst case. So balance and joint control must stay on the microcontroller.
- Robust to load. With the Linux CPU at 100%, the MCU dropped zero samples and typical latency rose only 0.4 ms.
- Found and fixed a hardware-level sensor glitch (an IMU register byte tear) and identified gimbal lock in the attitude filter, which drove the move to quaternions.
| Where | What |
|---|---|
| docs/ARCHITECTURE.md | The two-brain control design, and a log of design decisions with the evidence for each |
| docs/INNOVATION.md | The vision: an assistant that anticipates needs, with on-device, privacy-first learning |
| experiments/ | Subsystem experiments, each with code, raw data, analysis and conclusions |
| # | Experiment | Status |
|---|---|---|
| 01 | IMU Bridge Latency: real-time ↔ Linux link timing | ✅ Complete |
| 02 | Voice command pipeline: hear → decide → act | 🔜 Planned |
| 03 | Vision and stereo cameras for the head | 🔜 Planned |
| 04 | Battery state-of-health tracking | 🔜 Planned |
| 05 | CAN-FD actuator bring-up | 🔜 Planned |
Most assistant robots wait for a command. This one is meant to learn your routines and help before you ask, with all learning on-device so your data stays yours.
"The best assistant is one you don't have to ask."
Read the full innovation statement →
- Embedded and real-time: C++ on STM32, fixed-rate control loops, sensor fusion, register-level I²C debugging
- Systems: two-processor architecture, Bridge RPC interface design, clock synchronization, latency and jitter analysis
- Software and AI: Python, data analysis (pandas, NumPy, matplotlib), LLM-agent and continuous-learning system design
- Mechanical and electrical: CAD, carbon-fiber additive manufacturing, PCB design, battery and power-system design
- Method: hypothesis → measurement → published raw data → design decision
Follow the build at mjcaruana.com.