Menlo/asimov1-locomotion-26-10-05-upper-body

Asimov 1 locomotion policy checkpoint — 2026-10-05 independent upper-body control

• Velocity-commanded locomotion policy for Asimov 1, trained in Isaac Lab with PPO and adversarial motion priors (AMP).
• The policy controls the 12 leg joints. The waist and arms receive independent position targets, randomized during training so the policy learns to balance through upper-body motion.
• Exported from model_9500.pt in the 2026-10-05_17-24-12_upperik_new_ankle run.
• policy.onnx is the exported policy for inference, producing leg joint-position actions.
• policy.pt is the same policy as TorchScript, for running it from PyTorch.
• 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/cyclotron.

Independent upper-body control

The actor receives joint-position and joint-velocity feedback for all 23 motor joints, but outputs actions only for the 12 leg joints. A separate controller supplies position targets for the 11 waist and arm joints; those targets are not generated by this ONNX policy.

The previous-action observation contains the 12 raw leg actions followed by 11 zeros, retaining a 23-value history term and a total observation size of 78.

Ankle settings

• Ankle-roll action scale is 0.08; the other leg joints use 0.25.
• Ankle-roll position targets are clipped to [-0.09, 0.09] radians after applying the action scale and default pose.
• Ankle-pitch stiffness/damping are 224.42 / 10.201.
• Ankle-roll stiffness/damping are 35.2 / 1.6.

Use env.yaml for the full actuator settings, default pose, and action configuration.

How to run

Load policy.onnx with ONNX Runtime, or policy.pt with PyTorch, in an Asimov 1 runtime that supports 12 leg actions, independent upper-body position targets, and the observation recipe in env.yaml.

The runtime builds observations and runs the policy at 50 Hz. It converts each raw leg action into a position target using default_joint_pos + action_scale * action, then applies the ankle-roll target clipping. These operations happen outside the ONNX graph. Deployment metadata in the ONNX records the action joint order, scales, offsets, clips, and PD gains.

Inputs and outputs

• Input: float32 obs, shape [1, 78]: base angular velocity, projected gravity, velocity command (vx, vy, wz), full-body joint positions and velocities in motor-slot order, and the zero-padded previous action.
• Output: float32 actions, shape [1, 12], in this order: left hip pitch, hip roll, hip yaw, knee, ankle pitch, ankle roll; right hip pitch, hip roll, hip yaw, knee, ankle pitch, ankle roll.
• Control rate: 50 Hz (sim.dt = 0.005, decimation = 4).
• Trained in Isaac Lab (PhysX).

The full observation order and action joint list are recorded in env.yaml.