Menlo AI

Menlo Research builds Asimov, a developer-friendly humanoid robot aimed at researchers and engineers who want to run policies, collect data, and put on live demos on real hardware. With this launch, the team is open-sourcing the locomotion policy that ships with Asimov 1 along with the full training code behind it, so developers can start their next robotics project from a working baseline rather than from scratch. The release includes a trained policy checkpoint that runs in simulation and a training stack built on Isaac Lab. Beyond the checkpoint itself, the code exposes the pieces that are usually the hardest to reproduce, including hardware-tuned actuator configurations for the robot's motors and electronics, the reward definitions, system identification data, and domain randomization parameters. Training uses reinforcement learning combined with adversarial motion priors, where a style reward scores the policy against reference motion data, which is how the team gets natural-looking gaits and what lets users swap in a new reference file to change how the robot walks. Asimov 1 is 1.2m tall with 25 actuated degrees of freedom , and anyone modifying the hardware can encode the changes into URDF and retrain rather than relying on a policy tuned for a different build. Menlo is a global team spread across Singapore, Vietnam, and the US , and the team has flagged fault recovery from a fallen state as one direction they hope outside developers pick up first.