Asimov 1 Guide 2026: The $20,000 Open-Source Humanoid DIY Kit That Achieved Zero-Shot Sim2Real

---

Bottom line: Asimov 1 is the most open humanoid DIY kit — it proved that walking trained in simulation runs on the real robot with zero tuning

On August 27, 2026, the Asimov team (Menlo Research, Singapore) announced they achieved "zero-shot sim2real."

Asimov's official X post (@asimovinc, 177 likes, 32s video):

"Asimov is now zero-shot sim2real for locomotion. You train the policy in simulation. It runs on the robot without additional tuning."

In short:

• A walking policy trained in simulation runs on the physical robot with no additional tuning
• Blog: menlo.ai/research/zero-shot-sim2real-asimov
• Buy: menlo.ai/asimov-1 ($20,000)

Bottom line: the hardest problem in humanoid robotics — the sim-to-real gap — has been solved on an open-source DIY kit. One of the most important robotics announcements of 2026.

What is Asimov 1? Key facts

Why it matters: what is zero-shot sim2real?

The sim2real gap (the old problem)

In robotics, training AI in simulation is standard practice — experimenting on real robots is dangerous and slow.

But policies trained in simulation don't transfer to the real robot as-is — this is the long-standing "sim2real gap":

• Simulation has ideal physics (no friction/torque/latency variance)
• Real hardware has unit-to-unit variance, thermal drift, sensor noise
• The old workflow: train in simulation → spend hours-to-weeks tuning on the real robot

Menlo surveyed researchers: "How long before a new locomotion policy is live on your robot?" Answers ranged from "3-5 hours" to "2 weeks."

What Asimov achieved

Zero-shot sim2real = a policy trained in simulation runs on the real robot without additional tuning. Asimov closed this gap over 8 months.

The specifics (from the blog):

1. Zero-shot sim2real on flat-terrain walking — generalizes to small slopes and grass
2. Validated across multiple units — not one hand-tuned lab robot
3. Runs onboard at 50Hz — 25 motors + sensors, no external PC
4. Comparable to Unitree locomotion circa early 2025 — "still a long way to go"

Why it worked: hardware determinism

Menlo's key insight:

"The fix had less to do with nudging models than with making the underlying hardware a lot more deterministic."

• Minimize unit-to-unit variance (CNC 7075 aluminum, parallel actuation)
• RSU (Revolute-Spherical-Universal) ankle — two 36 Nm motors in parallel
• Design the robot as "a single deterministic hyperparameter in training"
• Don't overfit policies to hardware degradation (which trains overly conservative models)

Asimov 1 full specs

How to buy and build

Two buying options

1. Buy from Menlo Research (recommended, $20,000):

   • Bulk pricing on core hardware (industry-grade actuators cost $30,000+ bought individually)
   • Save ~$11,000 vs sourcing everything yourself
   • $499 refundable deposit to reserve your spot
   • The early material quoted a $15,000 target price; the final price when kits started shipping (August-September 2026) is $20,000 including shipping. The $15,000 figure in older articles is the goal, not the price (sources: the "Price goal" note on menlo.ai, and shipping-cost coverage on robottoday.com)
   • Ships in monthly batches

2. Source it yourself (BOM + CAD are free):

   • Buying actuators individually gets expensive
   • More time and effort, maximum learning

The build (100-hour path)

Asimov 1 is not plug-and-play:

• ~100 hours for mechanical assembly, wiring, flashing, calibration
• 3D-printed shells, CNC aluminum, harnesses, compute systems
• Basic walking policy included — advanced skills (dance, recovery) you train yourself

Try before you build

• Digital Asimov (https://try.menlo.ai/) — explore the robot in your browser, no hardware, no cost

Community

• Discord, forum, YouTube streams — "Build in Public"
• Small synchronized batches enable cohort-based troubleshooting

What's in the Kit: Included vs. Not Included

Here's a breakdown of what the Asimov 1 DIY Kit includes — and what it doesn't.

Important note: Asimov 1 doesnot include hands. The current configuration covers legs, torso, and arms — dexterous manipulation is waiting on future updates. Community-built hands are also an option.

---

What You Need for Assembly

Required Skills

Menlo Research is honest about this: "Assembling Asimov 1 is comparable in complexity to servicing a car. This is not a weekend project."

• Mechanical knowledge: Torque management for screws, understanding tolerances, handling CNC/3D-printed parts
• Electrical knowledge: CAN bus wiring, power management, motor driver configuration
• Software knowledge: ROS 2, Python, Linux (Raspberry Pi / Radxa setup)

If you've "done an oil change on your car at home," you can get started without issues. For a complete beginner, the bar is honestly quite high.

Required Tools

The kit doesn't include tools. You'll need to supply the following yourself:

• Hex key set (metric)
• Torque wrench (for tightening to specified torque values)
• Phillips screwdrivers (various sizes)
• Side cutters / wire strippers
• Multimeter (for electrical verification)
• Anti-static work mat (recommended)
• PC (Ubuntu recommended, for ROS 2 environment setup)

Assembly Time

50–100 hoursis the ballpark. The first milestone isn't "walking" — it's"getting to the point where you can safely power it on." From there, tuning and software setup take even more time before you can actually get it moving.

Menlo Research's recommended schedule:

1. Mechanical Assembly: Assemble parts, build the body with proper torque management (30–50 hours)
2. Electrical Wiring: Wire up CAN bus, power, sensors (10–20 hours)
3. Verification: Power-on checks, E-Stop operation verification, individual motor response tests (5–10 hours)
4. Software Setup: ROS 2, firmware, simulation environment (5–20 hours)

---

Setup Instructions (Software)

Once physical assembly is complete, it's time for software configuration.

1. Clone the Repository from GitHub

git clone https://github.com/asimovinc/asimov-1.git
cd asimov-1

2. Set Up the Simulation Environment

Asimov 1 supports MuJoCo-based physics simulation. You can train walking policies in simulation before running them on the actual robot.

# Simulation models are in the sim-model/ directory
cd sim-model
# MuJoCo setup (Python)
pip install mujoco
python -c "import mujoco; print('MuJoCo ready')"

3. Set Up the ROS 2 Environment

Asimov 1's control system is ROS 2-based. The following setup is required:

# Install ROS 2 Humble or the latest version
sudo apt install ros-humble-desktop

# Build the Asimov 1 ROS 2 packages
cd ~/ros2_ws
colcon build --packages-select asimov_bringup
source install/setup.bash

4. First Power-On

Important: Always perform the first power-onwith the E-Stop connected.

1. Connect the battery (or AC power)
2. Confirm that the Raspberry Pi 5 boots up
3. SSH into the Raspberry Pi: ssh pi@asimov.local
4. Check the motion control board (Radxa CM5) firmware
5. Verify CAN communication for each motor: cansend can0 000#
6. Manually test each joint one at a time

5. Apply Walking Policy

Asimov 1 ships with basic walking policies pre-configured:

# Start basic walking
ros2 launch asimov_bringup walk.launch.py

From here, you enter the phase of training your own walking policies with reinforcement learning, or tuning parameters to match your modifications.

---

The Full Scope of Open Source

The Asimov 1 GitHub repository (github.com/asimovinc/asimov-1) publicly hosts all of the following:

The BOM (Bill of Materials) is also publicly available, so you can source every part yourself and build from scratch: Asimov 1 BOM

Official documentation: docs.menlo.ai/asimov/1

---

The Menlo Research Ecosystem

The true value of Asimov 1 goes beyond the hardware. The entire ecosystem Menlo Research provides is what matters.

• Simulation Environment: MuJoCo-based — train walking policies
• Menlo Platform: The robot intelligence layer (under development)
• Intelligence Layer: LLM integration / AI agent orchestration
• Community: Discord, forums, batch system

As Asimov's CEO puts it, "Our bet is on the ecosystem." The goal isn't to build the best hardware — it's to provide a platform where the entire community can advance robot intelligence together.

---

Review

What's great

• Truly open source — BOM, CAD, simulation, policies all public
• Industry-lowest-class price — far below individual actuator sourcing ($30,000+)
• Zero-shot sim2real — dramatically faster research iteration
• Repairable — "A robot you can't repair isn't really yours"
• LLM integration — supports AI agent control research

Caveats

• $20,000 is expensive — for universities, labs, developers (not consumers)
• 100-hour build — requires skill and time
• No hands (0 fingers) — design/source your own end effectors
• Only basic walking included — advanced motion you train yourself (but deploys zero-shot)
• Rolling shipment — monthly small batches, wait times
• Final price varies by destination country, tariffs, shipping

Summary: Asimov 1 embodies the democratization of humanoid robotics

• Zero-shot sim2real achieved — simulated walking runs on the real robot as-is
• $20,000 DIY kit — industry-grade actuators at bulk prices
• Fully open source — BOM, CAD, simulator, policies
• 100-hour build — the robot itself is the lesson
• "A robot you can't repair isn't really yours" — breaking the black-box cycle

Instead of using a robot as a black box, open it, build it, train it, and make it yours — Asimov 1 delivers that philosophy at a realistic $20,000.

Questions Readers Ask

Q1. Does Asimov 1 arrive fully built?

No. It's a DIY kit requiring ~100 hours of assembly (mechanical, wiring, calibration). The process is the point.

Q2. Does a simulation-trained AI really work as-is?

Yes. That's the headline. Flat-terrain walking achieves zero-shot sim2real, validated across multiple units. Advanced skills still require your own training.

Q3. How do I buy it?

At menlo.ai/asimov-1. Reserve with a $499 refundable deposit, receive the kit for $20,000 (final price varies with shipping/tariffs).

Q4. Does it have hands?

No (0 fingers). You design/source your own end effectors.

Q5. Can I integrate an LLM?

Yes. The onboard Raspberry Pi 5 handles comms and high-level logic, enabling LLM-based agent control.

Q6. Can I buy it from Japan?

Yes, but final price varies with shipping and tariffs. Ships from Menlo Research (Singapore).

Related articles

• Quackd Guide 2026: giving a $399 robot an LLM brain
• OpenClaw 2.0 Guide 2026: the biggest AI agent update ever, explained
• DeepSeek-V4-Flash-Vision-Exp Guide 2026: DeepSeek's first multimodal AI agent