Tardigrade

I came across KT2 demo on Twitter and couldn't stop myself staring at it, felt so cool but sighed when I learned it's using scripted behaviours.

I wanted to do an action figure type open-source humanoid inspired by Asimov, because I don't have 20k to spare. And at the same time the idea clicked — why not build something simpler like KT2? Something like Asimov at action figure scale is more complex than it looks, so I decided to go for the KT2 form factor — it can do cool things while still helping me build the skillset. Although it might sound simple, there are grave challenges in deploying policies on such a tiny form factor.

My aim is to get a foundational understanding of layers in the physical AI stack. I want to rebuild it from scratch, understand every layer, and replace the scripted behaviours with something that actually learns — like how animals and insects figure things out on their own.

Why I'm Building This

This is the new fullstack. Every layer is something I either want to learn or get significantly better at:

• Actuators — very hyped to learn more and build from scratch, yeah firmware and low-level control included
• PCB design — for quite some time I've wanted to get my hands dirty on this
• CAD design — I've worked a bit with Onshape and liked it, now I want to improve it even further
• FEA and CFD — sounds cool to explore, always wanted to learn how they do this in aerodynamics and maybe we also make our Tardy learn to swim (would be fun to do structural analysis) :)
• System identification — I've partially done this but want to get a better idea of it
• Simulation — I've worked on this but there exists a plethora of things I still haven't figured out, particularly making USD and MJCF properly
• Learning — sim2real, world models, and downstream behaviours would be interesting to play with once I have foundations set up

Body Shell — 3D Model

Initial Motor Experiments

Architecture

Layer	Purpose
Electrical — Actuator	FOC driver PCB (KiCad)
Electrical — Mainboard	ESP32 main board
Mechanical — Actuator	Motor housing, gearbox CAD
Mechanical — Body	Frame, legs, feet
Firmware — STM32	FOC controller per actuator
Firmware — ESP32	Main control loop, comms, safety
Simulation — Model	MJCF + meshes
Simulation — Envs	Gymnasium environments
Training	RL training scripts + configs
Deployment	Policy → weights → MCU inference (C)

Key Design Decisions

Custom actuators over off-the-shelf servos. Torque and velocity control matter for learning-based policies. FOC-driven BLDC motors give direct torque control, multiple operating modes, and hopefully silent operation.

Learning over scripting. Scripted gaits are brittle. An animal doesn't have a lookup table for every terrain — it has a nervous system that adapts. The robot should build a repertoire of skills through interaction, not through my engineering effort.

Open source. This is the kind of project where sharing the work multiplies its value. Documenting everything as I go.

Status

Early-stage. Mechanical design underway in Onshape, project structure established, research into actuator design ongoing.

→ GitHub Repository