AsOverDog on Newton

This repository is a compact Newton/Kamino demo for asOverDog with Bennett, Planar, and Spherical legs.

Setup

Install uv and create the environment:

uv sync

Run

uv run play.py --robot bennett
uv run play.py --robot planar
uv run play.py --robot spherical

Viewer controls

Keys	Action
I / K	Forward / backward
J / L	Left / right
U / O	Turn left / right
P	Reset

Useful options

Run a fixed command without the viewer:

uv run play.py --robot bennett --headless --steps 200 --vx 0.5

Set fallback velocity commands for the viewer:

uv run play.py --robot planar --vx 0.5 --vy 0.0 --yaw-rate 0.0

Run the simulator with zero actions:

uv run play.py --robot spherical --zero-action

See all options with uv run play.py --help.

Models

Each robot pairs a USD mechanism with a robot-specific ONNX gait policy. All three policies use the same control interface and share one learned actuator model.

Robot	USD	Gait policy
Bennett	assets/robots/bennett.usdc	assets/policies/bennett.onnx
Planar	assets/robots/planar.usdc	assets/policies/planar.onnx
Spherical	assets/robots/spherical.usdc	assets/policies/spherical.onnx

At 50 Hz, each gait policy receives obs with 25 frames of 61 features (1 × 1525) and produces 12 normalized motor-position offsets through actions (1 × 12). The shared actuator model uses joint position-error and velocity history to convert those targets into motor torques at each simulation step. On CUDA, the policy, actuator, and eight 400 Hz Kamino physics steps remain on the device and replay as one policy-period CUDA Graph. Viewer updates are paced to the same 50 Hz rate; headless runs remain unthrottled.

Robot assets and simulation parameters are defined in assets/robots/robots.json. Policy paths, observation fields, tensor contracts, and hashes are defined in assets/policies/policies.json.

License

Licensed under the Apache License 2.0.