Genboo · 进步
Humanoid Cohort
ToddlerBot Education Partner
Cohort 03 · Now Reviewing Applications
Research-Informed Learning Model

Build, Simulate, and Train a Real Humanoid Robot. Mechanical · Electrical · Computer Science · AI.

Frontier humanoid-robotics research normally lives inside funded university labs — out of reach for the students most eager to do it. The Humanoid Cohort changes that: a selective 22-week research apprenticeship where high-school students reproduce and extend ToddlerBot, the open-source humanoid platform Stanford publishes on. Authorized and supported by the ToddlerBot team, with technical consultation from the ToddlerBot's original first two co-authors.

22
Weeks
30
Active DoFs
6 hr
Per Week
1:4
Mentor Ratio
Built for
High-school engineers

Aiming at top engineering programs. The cohort offers serious technical depth — a substantive research-style experience to draw on in applications and interviews.

Self-taught makers

Already shipping projects solo. Want mentorship from working researchers and access to a real lab setup instead of guessing alone.

Live · ToddlerBot
Capstone demo
Whole-body manipulation: pushing a loaded cart
Locomotion
RL · MJX trained
Perception
Stereo fisheye
Transfer
Sim → Real
What ToddlerBot Can Do

The platform's published research baseline.

Whole-body coordination, balanced contact forces, and sim-to-real RL transfer — demonstrated in the Stanford-published ToddlerBot paper. This video shows what the platform is capable of in research hands. Cohort 03 reproduces the platform and explores parts of this stack with mentor guidance — the exact deliverable each cohort ships depends on team progress and chosen track.

  • Reproduce, calibrate, and software-control the same humanoid platform used in published research.
  • Bring up MuJoCo simulation and run controlled motion experiments before applying AI.
  • Modify a baseline RL policy or motion experiment, analyze what changed, write it up, and present it.
The Platform

A research-grade humanoid.

ToddlerBot is the Stanford-published, open-source humanoid platform you'll build, wire, and bring up. Every component you work with is the same hardware used in active humanoid research labs.

ME
Mechanical Engineering

30 Active Degrees of Freedom

7 DoF per arm, 6 per leg, 2 in the neck, 2 in the waist. Ultra-robust 3D-printed structural components engineered for repeated drops, falls, and self-recovery.

  • Anthropomorphic 30-DoF kinematics
  • Dynamixel-class actuator integration
  • Print-in-place tolerancing & assembly
EE
Electrical Engineering

Jetson Orin NX Onboard Compute

Stereo fisheye perception, custom power management, and a clean serial bus topology — production-grade electrical design, not a hobby breadboard.

  • Stereo fisheye camera pipeline
  • Custom DC power distribution
  • Smart-bus actuator daisy-chain
CS / AI
Computer Science & AI

Sim-to-Real Reinforcement Learning

Build a faithful MuJoCo / MJX digital twin, identify system dynamics, and train RL walking policies that transfer to the real chassis.

  • MuJoCo + MJX physics simulation
  • System identification (sysID)
  • RL omnidirectional locomotion
Our Learning Model

A research apprenticeship, not a robotics class.

The cohort is designed around how learning and research actually work. Every design choice draws on established learning-science principles — mastery-based progression, learning by building, and deliberate practice with expert feedback — applied to a real, published research platform.

Constructionism

Learn by building real systems

Understanding is earned at the bench, not lectured. Students reproduce a research-grade humanoid from parts — the same platform Stanford publishes on — so every concept is anchored to something they built and can debug.

Mastery learning

Verify each layer before the next

One actuator → one limb → lower body → full chassis. Each stage has an explicit, checkable success criterion. No one advances on a layer they can't yet make work — difficulty is scaffolded, not dumped.

Responsible AI

Control before AI — no black boxes

Students reach reinforcement learning only after they understand the mechanics, electronics, and control underneath it. AI is a tool they can reason about and audit — not magic they invoke and trust blindly.

Research output

Ship paper, repo, and talk

Every student produces the artifacts research labs and admissions actually read: a short technical paper, an open-source contribution, and a recorded talk — concrete, measurable outcomes rather than a certificate of attendance.

Access & equity

Frontier humanoid research is normally locked inside well-funded labs. The cohort opens that work to high-school students, with need-based scholarships on hardware costs so cost isn't the barrier to entry.

Designed to scale

The cohort runs on a repeatable model: a research platform, a mastery-based curriculum, and a mentor ratio that holds. Each cohort's open-source contributions and documentation make the next cohort cheaper and faster to run.

Program Structure

Three phases. One robot. Built in pairs.

Sequential and cumulative. Each phase builds on the previous one. Schedule and intensity vary by phase — details below.

22 weeks total
Summer lab + Fall hybrid
1:4 mentor ratio

  Detail for each phase below

01
Phase 01 of 03 · Summer
Summer · 8 Weeks · ~32–48 contact hrs

Summer Intensive — Reproduce ToddlerBot

End-of-phase milestone

Reproduce ToddlerBot: from parts to a powered, calibrated, software-controllable humanoid platform. One actuator → one limb → lower body → full chassis. The Phase 1 success demo is not "the robot walks." It's "the robot powers on safely, every joint is recognized and calibrated, and it can run a simple verified motion sequence."

Powered, calibrated humanoid platform
Every joint verified & software-controllable
Engineering notebook + bring-up checklist
Weekly cadence
Summer Intensive 8 wk
Fall Academy 12 wk
Output Sprint 2 wk
Summer schedule · In-person, lab-heavy

Approximately 6 hours per week, weekend-weighted, in our lab space. Hands-on build sessions where pairs work side-by-side with mentors on physical assembly, wiring, and bring-up.

02
Phase 02 of 03 · Fall
Fall · 12 Weeks

Fall Academy — Simulate, Control & Improve

End-of-phase milestone

Move from joint-level scripts to controlled motion, then to simulation matching, then to learned-locomotion experiments. Each layer builds on the reproduced hardware platform from the Summer phase. Control before AI — students learn that AI sits on top of mechanics, electronics, and control, not in place of them.

Joint scripts, logging & teleop tools
MuJoCo model matched to your hardware
Controlled motion + RL experiments
Final project · pick your track

The final project splits into tracks so students can focus where they're strongest. Everyone learns the basics across the stack; the final deliverable goes deep on one of these directions:

  • Mechanical — improve cable routing or joint reliability on the chassis
  • Electronics — build a clean wiring map and bring-up checklist for future cohorts
  • Simulation — improve MuJoCo model matching to the real hardware
  • Control — create a stable keyframe motion (standing, squatting, stepping)
  • AI / RL — modify reward or observation in simulation, analyze the change
  • Documentation — build a student-friendly assembly guide from your engineering notebook
Weekly cadence
Summer Intensive 8 wk
Fall Academy 12 wk
Output Sprint 2 wk
Fall schedule · Hybrid — weekly virtual + in-person

Weeknight virtual seminars and code reviews, plus in-person lab integration days for sim-to-real testing on the robot you built in summer.

03
Phase 03 of 03 · Output Sprint
Dec – Jan · 2 Weeks

Output Sprint — Cohort Showcase

End-of-phase milestone

Turn your work into shareable artifacts. Write up your capstone in research-paper format, prepare an open-source contribution back to the ToddlerBot ecosystem where applicable, and present an 8-minute talk at the cohort showcase.

4-page technical paper (research-paper format)
Open-source PR to the ToddlerBot ecosystem
Demo reel + 8-min showcase talk
Why this matters

Most programs end at "I built it." This phase pushes students toward the artifacts that research labs and grad-school admissions actually read: a paper, a repo, and a talk. You leave with concrete deliverables you can share with researchers and mentors.

  • Reviewed by your mentor before the showcase
  • Help framing novel ideas into research-style writeups
  • Talk recorded for your application portfolio
  • Mentor introductions to the broader humanoid robotics community where relevant
Weekly cadence
Summer Intensive 8 wk
Fall Academy 12 wk
Output Sprint 2 wk
Sprint schedule · Daily virtual + showcase day

Daily virtual writing rooms and review sessions throughout the sprint, plus one in-person showcase day for talks and demos.

High-Level Syllabus

22 weeks, mapped to real engineering deliverables.

Each module is taught lab-first: a short concept primer, then hands-on time on the bench or in simulation, then a code review with your mentor.

Understand the robot as a system before you touch a single screw. ToddlerBot architecture, degrees of freedom, the actuator/sensor/controller/battery loop, lab safety, GitHub & engineering notebooks, and how to read CAD and assembly drawings.
Start with a single leg or arm module before tackling the full robot. One limb teaches almost everything: joints, servo orientation, fasteners, bearings, cable routing, alignment, torque limits, and debugging. Once one limb works cleanly, the rest is repetition — then assemble the full lower body.
Power system, Jetson, servo bus, IMU, connector labeling, multimeter basics, safe power-on sequence. Each joint gets its servo ID assigned and zero-points calibrated. Hard lesson: bad wiring beats good code every time.
Linux basics, Python environment, robot config files, simple joint test scripts. Phase 1 success demo: the robot powers on safely, every joint is recognized and calibrated, and it can run a simple verified motion sequence.
Git workflow, robot config files, joint IDs, logging, teleoperation scripts, basic sensor reading. Practical milestone: "Move joint 3 by 10 degrees, log commanded vs actual angle, and explain the difference." That is a real robotics lesson — not AI magic.
Why simulation matters. MuJoCo model and XML structure: joints, bodies, actuators, contact, gravity. The difference between simulation and the real robot. Match sim joint limits to your hardware. Simulation is a controlled approximation of reality — not a video game.
Open-loop vs closed-loop. Position control, PD control. Balance intuition. Gait as a sequence. Failure modes: slipping, tipping, delay, weak torque. Students create a simple standing, rocking, stepping, or keyframe motion. The lesson: AI sits on top of mechanics, electronics, and control — not in place of them.
Reinforcement learning fundamentals: policy, reward, observation, action. Domain randomization, the sim-to-real gap, and what can go wrong when deploying to hardware. Practical deliverable: run or modify a baseline policy or motion experiment, change one parameter, and analyze the behavior. Then iterate toward your final-project track.
Pick a track — Mechanical, Electronics, Simulation, Control, AI, or Documentation — and ship a portfolio-quality deliverable. Demo video, technical poster, GitHub documentation, short presentation, plus a written reflection: what failed, what changed, and what you learned.
Cohort 03 Promise

By the end of the cohort, students will understand the complete humanoid robotics pipeline — mechanical assembly, electronics bring-up, software control, simulation, and AI-based motion learning. Students work in pairs — every two students build one ToddlerBot platform together, and each student produces a portfolio-quality technical deliverable.

The Mentorship Advantage

You're not following a tutorial. You're working with researchers.

Every cohort student works on the same platform Stanford's research team publishes on. Code, simulation designs, and hardware decisions receive technical input from working researchers — including the ToddlerBot's original first two co-authors as Technical Consultants.

ToddlerBot team support
Authorized partner

Genboo is an authorized education partner of the ToddlerBot project. The ToddlerBot's original first two co-authors serve as our Technical Consultants — providing technical guidance on student code, simulation designs, and hardware builds throughout the cohort.

Genboo engineers as lead mentors
1:4 ratio

Working roboticists run weekly labs, code reviews, and 1:1s. A strict 1 mentor : 4 students ratio so your mentor knows your code, your robot's quirks, and your goals — not just your name.

Real open-source contributions

Your improvements to drivers, sim tooling, and policies can land upstream in the ToddlerBot repository — a portfolio piece that compounds.

Source platform
ToddlerBot

Stanford-published open-source humanoid research platform.

Active DoFs 30
Compute Jetson Orin NX
Perception Stereo fisheye
Sim MuJoCo + MJX
Policy RL omnidirectional
Apply · Cohort 03

Admission is selective.

Cohort 03 admits a small group each cycle — hardware and lab capacity are the binding constraints, so we review applications carefully and look for students who will get the most from a research apprenticeship and help carry the work forward. Prior programming exposure (Python / C++) helps, but intellectual curiosity and problem-solving drive matter more than credentials. Participants leave with a working humanoid platform, research-grade artifacts, and a mentor network in humanoid robotics.

Open to current high school students (grades 9–12).
Applications are reviewed on a rolling basis; we encourage applying early.
Need-based scholarships available for hardware kit costs.
Costs discussed once shortlisted

Shortlisted applicants discuss program fees, hardware kit costs, and scholarship eligibility during the interview. We're committed to keeping the cohort financially accessible to admitted students regardless of means.

Express interest

Start your application

We'll follow up within 5 business days with next steps.