chris thomas
mechatronics & applied robot learning
— k-botopen source · CT-2025
▶ toasting bread

— team

A small crew punching way above our weight: Benjamin Bolte, Aaron Xie, Ethan Fraser, Ian Dyke, Joshua Huang, Jaffer Ali, Paweł Budzianowski, JX Mo, Jan Schmidt, Wesley Maa, Ali Kuwajerwala, Rishabh "Greg", Scott Carlson.

— as seen on

the information ↗ · hacker news ↗

— intro

Over winter break, I started a skunkworks project within k-scale with two other engineers to design an opinionated humanoid robot. It became the k-bot: the first open-source humanoid platform capable of reliably deploying RL locomotion policies and VLA for manipulation.

I owned the hardware stack from mechanical design through sim2real deployment, ensuring the hardware served the intelligence stack. I led prototyping, validation, system integration, and manufacturing, bringing this robot to life in less than 4 months.

▶ boxing match
mp trials
▶ rl walking deployment
▶ rl sim walking
a proud parent

— outcome

A humanoid robot development platform focused on machine learning applications in locomotion and manipulation. Affordable with a BOM under $13k, repairable using COTS parts, design optimized for assembly: highly modular with strong scaling capability to adapt to any task.

150+ preorders for $2M. Began mass production trials with 35 units deployed by developers and researchers worldwide.

— devlog

December 17th, 3am. Analytical teardown of our previous platforms and competitors like unitree and booster, documenting failure modes, repair blockers, and the cost of proprietary hardware. Turned those findings into a PRD covering mechanical, electrical, controls, and ML architecture, then sketched out features, mechanisms, and form.

v1: stompy
stompy teardown
unitree teardown
booster and stompy

First CAD pass was a rough block of the full robot, laid out spatially to validate architecture. Second pass went deeper: every component designed for 3D printing, assembly, and testing. Minimum viable strength, functional enough to test reliably. Ran FEA to validate load paths, printed everything, assembled, and tested end to end.

leg wiring harness sketch
arm wiring harness sketch
hardstops, wiring and assembly
constraining shoulder wires
form and scale
shoulder design sketches

Third iteration locked the design. Full robot assembled in 3D printed parts, verified, then adjusted for CNC machining. Two months from conception, we had a full CNC prototype running. Ran a DFX tracker through assembly, testing, and deployment, logging every electrical, thermal, and mechanical issue.

printed torso assembly
upper body design
team validating assembly
final iteration CAD
printed leg assembly
head design and assembly
first CNC batch
lightweighting
hip wiring validation
torso and leg CAD
shoulder and arm CAD
large action space and ROM

Fourth iteration was production-ready. 35 units manufactured. With hardware locked, we closed the sim2real gap: locomotion policies trained in simulation deployed successfully on the physical robot. Full-body RL walking and VLA manipulation, on a platform built from scratch in under four months.

k-scale team with the k-bot
mass production trials
k-bot modularity demo
k-bot walking on grass
candid production with the k-bot

— reflections

Most of what I'd change comes down to spending more time before committing to decisions that are hard to reverse, and more time after shipping to learn from how people actually use the robot.

More iterations: two or three more cycles of prototyping, developer testing, and feedback.
More upfront research into actuator selection, limb proportions, and morphology before entering CAD.
Better understanding of the end-use action space: what tasks matter, what capabilities are worth the extra effort.
Standardized modular interfaces, connectors, compute and power blocks for repairability and cost.
Design for human interaction: soft contact surfaces, safe collision behavior, capacity for expression.
next: z-bot v3 →