Asimov 1 locomotion policy checkpoint

• Basic velocity-commanded locomotion policy for Asimov 1, trained in Isaac Lab with PPO and adversarial motion priors (AMP).
• policy.onnx is the exported policy for inference, producing joint-position actions.
• agent.yaml records the actor-critic and AMP training settings, including the motion data configuration.
• env.yaml records the simulation and locomotion task settings, including observations, actions, commands, and rewards.
• Training and evaluation code: menloresearch/isaac_asimov.

How to run

Run the policy in a browser simulation of Asimov 1 with humanoid-policy-viewer. It needs Node, git and bash, and no GPU:

git clone https://github.com/menloresearch/humanoid-policy-viewer
cd humanoid-policy-viewer
npm run hf Menlo/asimov1-locomotion-0818

This downloads the policy, starts the viewer and opens it in your browser. Drive the robot with the velocity sliders, push it, or run the benchmark suite.

Inputs and outputs

• Input: 78 values at 50 Hz: base angular velocity, projected gravity, velocity command (vx, vy, wz), joint positions and velocities, and the previous action. All of these are available on the real robot.
• Output: 23 joint position actions (legs, waist and arms; the neck is not driven). Each target is default_joint_pos + action_scale * action, with both values in env.yaml.
• Trained in Isaac Lab (PhysX); the viewer simulates it in MuJoCo.

The full observation order and joint list are in the viewer's supported policies doc.