dene33/asimov1-forward-jump

Asimov 1 forward jump

A forward jump on the walking policy's velocity command, trained in Isaac Lab with PPO from a retargeted human jump. Validated in simulation on 18 October 2026: the ONNX matches all 14 deployment contract fields of Menlo/asimov1-locomotion-26-09-24; in Isaac Sim 0 of 1,024 robots fell nominally, 1 with the training randomisation (before the jump command) and 7 with pushes; in MuJoCo, on the published model with the trained joint armature and the firmware's ankle limits, the pelvis rose 12.9 cm and the robot travelled 40 cm, both feet leaving the ground within 20 ms of each other. It turns about 14 degrees to its right over the jump and lands in a slightly deeper crouch than the reference. Not tested on a physical robot.

A forward jump on command for Asimov 1, with the walking policy's interface: it runs where the walking policy runs.

Interface

policy.onnx has the deployment contract of the walking policy it was trained against (Menlo/asimov1-locomotion-26-09-24): the same entries in its ONNX metadata for the observations, joint order, action scales, offsets and clips, PD gains and 50 Hz rate. A runtime that runs that walking policy runs this one with the same settings:

• 78 observations: the pelvis angular velocity x 0.25, gravity's direction, the velocity command (forward speed, sideways speed, yaw rate), the joint positions relative to the default pose and the joint velocities x 0.1 in the motor-slot order, and the previous action.
• 23 joint position targets: the default pose plus the action times its scale (0.08 for the ankle rolls, 0.25 for the other joints), with the ankle rolls clipped to 0.09 rad either way.

The velocity command is the trigger:

• Zero: the robot stands and balances.
• A forward speed of 0.3 m/s (it was trained with 0.25 to 0.4), held for the whole jump, about 2.2 s, then released: one forward jump. The reference it learned from lands 1.4 s after the command starts; release the command after landing. Held on after that, it jumps again.

It doesn't walk: hand over to the walking policy once it stands.

Files

• policy.onnx: the policy, from obs (1, 78) to actions (1, 23), with its observation normalisation and the deployment contract in its metadata.
• env.yaml and agent.yaml: the training task and PPO settings, including the observation recipe, joint order, default pose, PD gains and effort limits.
• policy.pt is the same policy as TorchScript, for running it from PyTorch.
• code_state.yaml records the code the policy was trained with: the commit, the robot model, the package versions, a hash of the network's code, the joints and gains the policy's inputs and outputs resolved to, and (for runs that record them) the sha256 of the uploaded env.yaml and agent.yaml.

How it was trained

Asimov1-Jump-v0 in the training code, with PPO in Isaac Lab: the motion tracking rewards of Asimov1-Tracking-v0 on cmu_13_11_forward_jump_froude_nolift.npz, plus the legs' left/right symmetry and lifting and landing together, the arms' and waist's joint angles, the heading and smoothness. The policy never sees the reference motion: it only shapes the rewards and what the critic sees, so the policy doesn't need it to run. The ankle targets are clamped in training as Asimov 1's firmware clamps them (0.35 rad pitch, 0.1 rad roll), and the joints' armature is randomised.

Reference motion: CMU Graphics Lab Motion Capture Database (mocap.cs.cmu.edu), subject 13, motion 11, retargeted to Asimov 1. The database was created with funding from NSF EIA-0196217.

Status

Checkpoint model_5498.pt. Trained in simulation only. Before a robot, play it in MuJoCo (./cyclotron.sh --sim2sim), then on the robot with ./cyclotron.sh --sim2real, hanging from its gantry hook.

Training and evaluation code: Dene33/cyclotron/tree/motion-tracking.