Asimov open-sources the Isaac Lab training stack behind its $15,000 DIY humanoid’s walk

Asimov, the humanoid project of Singapore-based Menlo Research, on September 25, 2026 published the code it uses to train its Asimov 1 robot to walk. The isaac_asimov repository is a standalone NVIDIA Isaac Lab extension under the BSD-3-Clause licence, and trains velocity-tracking walking policies with PPO and adversarial motion priors (AMP).

Why it matters: most humanoid makers treat the locomotion controller as the crown jewel. Asimov now gives buyers of a $15,000 kit the full recipe to retrain it, which makes its robot a research platform rather than a closed product.

What did Asimov open-source?

Asimov released the training and evaluation code for the Asimov 1 walking controller, not just a robot model. According to Humanoids Daily, the release covers the Isaac Lab training setup, reward definitions, actuator settings and the simulation variations used to make the controller survive the jump to hardware. The README pins specific versions of Isaac Lab and an asimov-1 robot description as git submodules and installs them with a single script on Ubuntu 22.04 or later.

A quick test run of 128 environments and 100 iterations takes about ten minutes on an RTX 4090, the README says. The full baseline uses 4,096 environments on a single A6000 or RTX PRO 6000, and the code supports multi-GPU runs through PyTorch distributed training. Trained policies can be exported to ONNX for deployment. Asimov describes the release as “a starting point for adapting the controller to hardware changes, exploring different walking styles and developing further capabilities.”

What are adversarial motion priors?

AMP adds a second learning signal to standard reinforcement learning: a discriminator scores how closely the robot’s motion resembles reference clips, and the policy is rewarded for looking natural as well as for tracking the commanded speed. Asimov recommends the AMP task; the plain PPO task is kept as a baseline. The repository credits beyondAMP from Renforce Dynamics for its AMP implementation and cites HybridRobotics’ whole_body_tracking and MuJoCo-based mjlab as references. Our explainer on how sim-to-real transfer works covers why randomised actuator and terrain settings matter as much as the reward itself.

The practical point is that rewards and actuator parameters are where most of the tuning time in legged locomotion goes. Publishing them lets a builder who swaps a motor or changes a leg length retrain rather than start from zero.

Is the trained Asimov checkpoint available?

No. The announcement said a trained policy checkpoint was included, but Humanoids Daily could not find a downloadable checkpoint in the repository or its GitHub releases, and we could not find one either as of September 28. The README’s “play” command loads the latest checkpoint from a user’s own training logs. Until the weights appear, owners must train the policy themselves on a high-end NVIDIA GPU.

The repository was small at the time of writing: seven commits and 15 GitHub stars. Asimov says it plans community livestreams where developer-contributed policies are tested on a real Asimov 1.

Who makes the Asimov humanoid?

Menlo Research, registered in Singapore, launched the Asimov 1 “Here Be Dragons” DIY kit on March 3, 2026 at a target price of $15,000, set at bill-of-materials cost, with a $499 deposit. The robot stands 1.2 m and weighs 35 kg. Most structural parts are designed for Multi Jet Fusion 3D printing, and it uses an RSU ankle mechanism and passive toes. On April 27 the company published the mechanical CAD, wiring and a MuJoCo simulation model under CERN-OHL-S-2.0 for hardware and GPL-2.0 for software. Humanoids Daily reports the team has already shipped DIY kits and puts assembly at roughly 50 to 100 hours.

The release fits a wider pattern of open robot software, from Intrinsic’s Apache 2.0 robot runtime and NVIDIA’s Isaac ROS 5.0 to Unitree’s 6B-parameter UnifoLM-WLA-1.0 model. The difference is scale: Asimov is opening the full stack of a small kit robot, while larger companies open selected models and software layers and keep their hardware closed.