Asimov 1, end to end

The Asimov Workshop is an independent teaching app, written by Sergej Stein and not affiliated with Menlo Research, about Menlo’s open-source Asimov 1 humanoid. It runs in the browser (three.js, with MuJoCo 3.3.8 compiled to WebAssembly). BUILD assembles any of six modules step by step from the published CAD, 1,184 parts in all. UNDERSTAND explodes and sections the robot, compares the hardware CAD with the simulation model and sweeps the joints, including the parallel ankle. TEST runs Menlo’s published walking policy live. WATCH is a one-minute tour. This page does not copy any of it. It cites numbers the workshop publishes and sends you there for the 3D.

The robot, per Menlo’s documentation, has 25 powered joints, stands about 1.2 m and weighs about 35 kg, with a Radxa CM5 for motion control and a Raspberry Pi 5 for media and networking. The workshop’s model locks the neck, so the simulation has 23 actuated joints. Menlo’s SDK and firmware pages are marked as in development.

The walking policy is a contract. A 78-value observation goes in: angular velocity, projected gravity, the velocity command, joint positions and velocities, and the previous actions. 23 actions come out, scaled by 0.25 around a default pose. The policy runs at 50 Hz while physics steps every 5 ms, four physics steps to one decision. Training covered forward commands from -0.6 to 0.8 m/s.

Below the policy sits ordinary control. Each action becomes a joint target for a PD loop, with stiffness 150 on the hips and knees and damping capped at 5, a torque limit per joint (45 N·m on the hip pitch, hip roll and knee, lower on others), and an actuator behind it with a rated and a peak torque: for instance the hip pitch, an EC-A6416, is 40 N·m rated and 120 peak, and the knee, an EC-A4315, 25 rated and 75 peak.

The ankle is a parallel mechanism: two motors drive it through push rods, an RSU linkage. The firmware turns pitch and roll into the two crank angles with a linear map, A = 2.02 pitch - 0.8 roll and B = -2.02 pitch - 0.8 roll. Against the exact linkage closure that map is off by up to 0.043 rad in pitch and 0.039 rad in roll at the corners of the range, which is ±0.35 rad in pitch and ±0.10 rad in roll. The workshop’s ankle view shows the real geometry.

Then the caveats, which the workshop publishes itself. The policy was trained in Isaac Lab, on PhysX, and runs here in MuJoCo, so what you see is a sim-to-sim check, not a test on the robot; the workshop’s own README says its simulation results say nothing direct about the physical robot. The model also disagrees with the hardware in places: it weighs 32.22 kg against the published 35, the thigh and shank lengths differ between the CAD and the model, the knee axis sits 23.6 mm higher in the CAD, and foot friction is 0.6 in the model while training randomised it from 0.8 to 1.2.

The robot adds behaviour the simulation lacks. Its firmware caps speed at 0.4 m/s while training covered up to 0.8, so faster commands are simulation only. On the robot a fall latches a limp DAMP state until the firmware restarts. The workshop’s viewer, instead, holds the robot softly after a fall (kp 10, kd 3), and counts a fall at about 80 degrees; those are the viewer’s behaviour, not the robot’s.

The practice is as useful as the robot. The workshop labels every value as source spec, sim value, CAD derived, approximation or illustrative, pins every input file to a hash, and publishes the disagreements in a discrepancies file. That is the same discipline as this site’s software-defined-vehicle pages, and the one to copy.