To scale down costs while keeping the software architecture (state machines,
ring buffers, multi-node communication) identical to your final build, the key
is to mimic the behavior of expensive QDD motors and industrial buses using
hobby-grade hardware.
Here is a highly economical parts list for a mini-scale MVP that perfectly
preserves your modular "transformer" architecture:
1. Actuation: Smart Serial Servos
Standard PWM RC servos are cheap but useless for this because they lack
telemetry (position, velocity, and load feedback) and run on a star topology.
You need actuators that mimic the API of high-end robotic joints.
 * Feetech STS3215 (or SCS15): (~$15$20 each) These are the gold standard for
budget robotics. They are daisy-chainable serial bus servos that provide
real-time feedback on position, load, speed, and temperature. They support
both continuous rotation (velocity mode for the rover wheels) and position
control (for the humanoid joints), perfectly matching your Stage 1 to Stage 2
transition requirements.
 * Dynamixel XL330-M288-T: (~$25 each) Slightly more expensive, but the
Dynamixel SDK is industry standard and heavily documented for ROS2/custom
stacks.
2. Compute & Communications Bus
Since you are already working with the RP2040 (from your MIDI controller
work), you can use it to build out the distributed node architecture very
cheaply.
 * Master Torso Controller: A Raspberry Pi Zero 2 W or Orange Pi Zero 3
(~$15$25). They run a full Debian environment, which fits your terminal-first
workflow, and are powerful enough to run your C++/Node.js state machines,
kinematics math, and local LLM routing calls over Wi-Fi.
 * The Bus (RS485 vs CAN FD): Smart servos use a TTL or RS485 serial bus
instead of CAN FD. To keep the code structure similar, use an RP2040 (Pi Pico)
(~$4) as your sub-controller. It translates your torso's commands into the
servo's serial protocol.
 * If CAN FD is strictly required for testing: Buy cheap MCP2562FD CAN
transceivers (~$2) and wire them to Pi Picos. Each Pico acts as a CAN node at
the socket, interpreting CAN frames and driving a cheap generic DC motor or
servo.
3. Power Distribution
At mini-scale, you avoid the danger and expense of 48V systems.
 * Battery: A standard RC 2S (7.4V) or 3S (11.1V) LiPo battery (~$15). Serial
servos are designed to run directly off these voltages.
 * Logic Power: A mini MP1584EN buck converter (~$1) to step the LiPo voltage
down to a clean 5V for the Pi Zero and RP2040s.
4. Chassis & Modularity
 * Frame: Pure 3D-printed PETG or PLA. Design the torso as a hollow box with
standardized bolt patterns (like a 20mm grid) at the four corners.
 * Wheels: Generic 1/16 scale RC car wheels with a 12mm hex hub (~$10 for a
set of 4). You can 3D print an adapter to mount the hex hub directly to the
spline of the serial servo.
 * Wiring: Standard 3-pin JST-XH cables to daisy-chain power and data from
socket to socket.
The Economic Advantage: For under $150, you get a fully functional, 4-node
networked robot. Because you are using bus-driven servos with
velocity/position modes, the code you write to query joint states and command
wheel velocities over the Pi Pico will port almost 1:1 to your final
high-current CAN FD architecture.