Your first humanoid RL robot
Master walking, balancing and locomotion on a desk-sized humanoid
Pave your way for future human-scale robots
Master walking, balancing and locomotion on a desk-sized humanoid. Build skills now, then scale to human-size robots.
Pre-order open · Ships Q4 2026
APPROVED
TRANSFER WORKFLOW
You can train RL on this 50 cm desktop robot and transfer everything you learn seamlessly to human-scale robots, since they share similar:
| What transfers | Details | Status |
|---|---|---|
| Core RL formulation | Proprioceptive/joint-state inputs → Neural policy → Joint-position commands | ✓ |
| Simulation platforms | Nvidia gym, MuJoCo | ✓ |
| Core algorithm | PPO — similar algorithm used in human-size humanoid locomotion research | ✓ |
| Humanoid leg structure | 12-DoF bipedal configuration matching the larger humanoid platform | ✓ |
| Transferable domain | Joint sensing, IMU-based state/control logic, reward design, and RL workflow all carry over | ✓ |
WHY HUMANOID
Every other platform forces you to compromise. M1 is the only desk-sized robot that gives you real bipedal walking + manipulation + human structural alignment.
The Proof
M1 isn't just "smaller" — it uses the similar fundamental techniques as the big guys. Here's why what you train on M1 matters.
01
Train in your environment of choice — NVIDIA Isaac Gym and MuJoCo with our Python API.
02
The IMU + joint angle + joint velocity control loop in M1 is identical to what full-size humanoids use. Walking, balancing, posture correction — the algorithm architecture is the similar.
03
Major humanoid company trains with PPO. M1 uses the similar algorithm, the similar reward function patterns, and the similar training loop.
04
M1 has a similar DOF structure to full-size humanoid legs — 6 DOF per leg (hip yaw/roll/pitch, knee pitch, ankle roll/pitch).
05
Onboard camera with real-time face detection and verification — just like full-size humanoid vision stacks. Train perception pipelines, identity recognition, and human-robot interaction skills directly on M1.
06
M1 navigates real-world environments with obstacle detection and path planning — a similar navigation stack architecture to full-size humanoids. Deploy SLAM, dynamic obstacle avoidance, and waypoint navigation at desk scale.
PROOF 01
Simulation Platform
Sim → Real Transfer
Simulation Match
Proof 02
Sensors & Control
Proof 03
PPO Algorithm
Proof 04
12-DOF Bipedal Legs
Proof 05
Face Recognition
Proof 06
Navigation & Obstacle Avoidance
Get Started
From unboxing to your first trained walking policy — everything you need, in one place.
01
Train
Starter kit with training assets and real-robot deployment files. Follow the docs, set up your local env, and run simulation — all in one repo.
02
Deploy
Python-native deployment with a one-line terminal command. Or hit LB + X on the controller — deploy in two ways, full source code included.
03
Control
Walk, turn, and trigger actions with the bundled controller. Plug in and play — no configuration needed.
04
Support
Stuck on setup, training, or deployment? Our support team is one email away — we usually reply within 24 hours.
Watch M1 go from simulation training to real-world walking.
Watch demo
Small details that make a big difference in your daily hacking.
Why You'll Love It
The face display interface is fully customizable — code your own expressions, moods, and status indicators.
Approved
Multiple Modal
Swap to the included Cart for 1.1 m/s speed, navigation and SLAM research.
Free Cart — $200 value, included
Safety
At just 50cm tall and 2.8kg with servo actuators, M1 is inherently safe. Train, test, and tinker without worry.
Training
No cluster needed. M1's lightweight sim + PPO pipeline runs on a consumer GPU laptop. Iterate fast on your own hardware.
Consumer GPU — no cluster required
Own It
No more waiting for lab time. M1 fits on your desk, runs on your laptop, and it's fully yours.
Full access — hack it, break it, own it
Anywhere Training
Full-size humanoids need safety gantries, harnesses, and dedicated lab space just to get started. M1 fits on your desk — train, test, and iterate in any room, office, or classroom without specialized infrastructure.
Desk-ready — zero infrastructure needed
Quiet Operation
M1's servo-driven design runs quieter, softer than normal office conversation. Train all day without disturbing anyone around you.
Quieter than normal office conversation
Built-in Tasks
Six reference behaviors pre-loaded on M1. Train from scratch or fine-tune them — your starting point doesn't have to be zero.
Fast & slow gait. Speed-adaptive walking on flat ground.
Rough ground, slopes & uneven surfaces. Adaptive balance.
Single-leg forward kick. Dynamic balance on one foot.
Self-righting recovery. Stand up from a fall.
Full-body choreography. Rhythmic motion control.
Stand & walk on a moving platform. Active stabilization.
BUILT FOR DEVELOPERS
Designed for PyTorch/ONNX model export and sim-to-real pipelines via the Reinforced Imitation Framework.
Open SDK with full-stack control — gain direct access at every layer with the Python SDK or the M1 Spine SDK for low-latency control. Built for ROS 2 integration and Linux + OTA workflows.
PyTorch/ONNX sim-to-real pipelines, open SDK, ROS 2, Linux + OTA, with Python and M1 Spine SDK control.
M1 / SIDE SYSTEM VIEW
TRAIN SMALL. SCALE SMART.
Master the fundamentals on M1, then take your skills to any full-size humanoid platform. Don't scale the cost before you scale the skills.
M1
Desk-sized bipedal research
Mid-Size Research
Lightweight research
Full-Size Humanoid
Consumer full-size
Industrial Grade
Industrial full-size
Heights and prices are approximate, for comparison only.
UNBOXING
Everything you need to start training — right out of the case. Portable, protected, and ready to go.
Eight essentials packed in a custom EPP foam carry case. Everything you need to get started — just unbox and train.
Join the early-bird list and be among the first to get your M1. Only 100 units for the first batch in Novbember.
Discord Community
Join our tech support community for live help, tutorials, and fellow M1 builders.
Join Discord →
Email Support
Send us a message — we usually reply within 24 hours on business days.
FAQ
Everything you need to know about ordering, shipping, warranty, and development with M1.
Yes — full refund any time before dispatch.
At this time, we ship to the United States only.
The included 1-year warranty is also exclusive to U.S. customers. This provision applies to the extent permitted by applicable law. If your local consumer protection laws grant you more favorable rights (e.g., a statutory minimum warranty period), this provision does not limit or waive those rights. For full details on your statutory rights, please consult your local consumer protection regulations.
Yes. M1 EDU Combo provides reference workflows designed for both Isaac Gym and MuJoCo. Policies trained in PyTorch can be easily exported via ONNX and deployed to the physical robot through the Reinforced Imitation Framework for efficient sim-to-real transfer. Example training scripts and reference policies are available directly through the private developer Discord community.
To deliver maximum flexibility for development and control, the M1 EDU Combo SDK features a robust Python API for high-level behavioral workflows alongside the low-latency Micro Robot Spine SDK for direct motor control commands. M1 EDU Combo is architected for seamless integration with ROS 2 environments at the message-bus level.
Yes. M1 EDU Combo features an integrated 10 TOPS NPU (BPU) optimized for local neural network inference, processing real-time spatial data directly from the dual binocular cameras. Developers can deploy custom models via the standard ONNX toolchain, and the Zeroth engineering team is available to assist with model quantization recipes and operator optimization.
Zeorth provides dedicated support to keep your hardware operational. For any maintenance, chassis servicing, or component repair inquiries, developers can contact the Zeroth support team directly for guided technical assistance and service options.
To safeguard the robot's high-torque actuators during dynamic policy testing, M1 EDU Combo features built-in hardware force limits by default. This ensures safe operation and long-term durability during aggressive movement, push recovery, and contact-rich reinforcement learning experiments.
