# Roboto Origin model

Source: https://github.com/Roboparty/rpo_description
Commit: 37aac9ca665e92731444a1618320078e7ba21569

Unmodified mjcf/rpo.xml and meshes/*.STL, licensed under CERN-OHL-W-2.0.
See LICENSE. The browser loads rpo.mjb, precompiled with MuJoCo 3.15.0 from
simplified meshes in lite/. Inertias, joints and original convex mesh collision
settings are retained. Unused textures and heightfields are removed.
Reproduce with npm ci && npm run prepare:sim && npm run build.
Integration sources are served at /sim/source/main.js, /sim/source/policy.js,
/sim/source/build-sim.mjs, /sim/source/prepare-sim.mjs, /sim/source/sim.html,
/sim/source/sim.css, /sim/source/loading.js, /sim/source/i18n-labels.html,
/sim/source/en.toml, /sim/source/zh.toml, /sim/source/package.json,
/sim/source/package-lock.json, /sim/source/playback.mjs, /sim/source/npy.mjs,
/sim/source/policy-runtime.mjs, /sim/source/catalog.mjs, /sim/source/render.mjs,
/sim/source/import-sim-motions.mjs, /sim/source/import-sim-locomotion.mjs.
The shared build cache helper is served at /sim/source/build-cache.mjs.
The policy integration source is distributed under GPL-3.0 (POLICY-LICENSE).

## RP1 model and policies

Source: https://github.com/Roboparty/RP1_description
Commit: c2956c58b7381c0e7433e200ef59930468fe534b
File: mjcf/robot_24dof.xml and the referenced meshes.

Source: https://github.com/Roboparty/roboparty_deploy
Inference submodule commit: 7a1178256d2b0f0e27427248076cfadc4db12f4d
Files: `src/inference/robots/rp1_24dof/models/policy.onnx`,
`policy_amp.onnx`, `policy_ufo.onnx`, `policy_wave.onnx`,
and the referenced motion and latent sequences.

The RP1 browser assets are distributed under their upstream terms. The
compiled model uses simplified visual meshes and MuJoCo 3.15.0; see
`rp1/SOURCE.md` for the model-specific attribution.

## RPO AMP policy

Deployment source: https://github.com/Roboparty/roboparty_deploy
Commit: e371ad6d9250b3b63f22fcfdee47b40c485b28ac

Inference source: https://github.com/Roboparty/roboparty_inference
Commit: 7a1178256d2b0f0e27427248076cfadc4db12f4d
Files: `robots/rpo/models/policy_amp.onnx` and `robots/rpo/configs/amp.yaml`.
The ONNX model is unmodified and served as `rpo/policy_amp.onnx`.
The original configuration snapshot is served as `rpo/amp.yaml`.

AMP uses 78 proprioceptive observations, three observation-major history
frames (234 inputs), and 23 actions at 50 Hz. Its command bounds are
vx [-0.5, 2.0] m/s, vy [-0.6, 0.6] m/s, and yaw [-1.57, 1.57] rad/s.

License: GPL-3.0, see POLICY-LICENSE.

Reproduce these imports:

```
git clone https://github.com/Roboparty/roboparty_deploy /tmp/roboparty_deploy
git -C /tmp/roboparty_deploy checkout e371ad6d9250b3b63f22fcfdee47b40c485b28ac
git -C /tmp/roboparty_deploy submodule update --init src/inference
node scripts/import-sim-locomotion.mjs /tmp/roboparty_deploy
```

## BeyondMimic motion policies

Source: https://github.com/Roboparty/roboparty_inference
Commit: 7a1178256d2b0f0e27427248076cfadc4db12f4d
Deployment repository: https://github.com/Roboparty/roboparty_deploy

RPO assets are copied from `robots/rpo/models/policy_wave.onnx`,
`policy_dance0.onnx`, `policy_dance1.onnx`, and `policy_getup.onnx`, paired
with their corresponding files in `robots/rpo/motions/`. RP1 assets use
`robots/rp1_24dof/models/policy_wave.onnx` and `motions/wave.npz`.
The ONNX models are unmodified. The `joint_pos` and `joint_vel` NPZ arrays
are concatenated as little-endian float32 NPY files, preserving the upstream
policy joint order and 50 FPS sampling. Per-motion JSON metadata records the
frame count, FPS, and joint count. Where body transforms exist, the initial
root position and quaternion are taken from frame 0, body 0 (`base_link`),
with quaternion order `w, x, y, z`. These transforms are used exclusively
to initialize the scene and are not additional policy inputs. Body 0 and
the quaternion convention were verified against RPO MuJoCo forward kinematics.
RP1 Wave contains only joint arrays and FPS, so its metadata has no root pose.
Other NPZ fields are omitted.

Motion policies initialize the robot at the first reference joint pose and,
where available, the first root pose, with physics paused. Starting physics
without triggering the motion holds frame 0's reference joint positions with
zero reference velocities; triggering playback enables the original reference
positions and velocities. UFO latent files contain only `z`, without joint or
root poses, so UFO initializes at the deployment default pose with latent
frame 0 selected. A locomotion policy does not run while waiting.
Playback uses a separate trigger from the physics simulation. Selecting a
sequence loads only its matching model and data; simulation pause also pauses
sequence advancement. Each sequence is played once and retains its last
reference frame until stopped or restarted.

Reproduce these imports with Python 3 and NumPy installed:

```
git clone https://github.com/Roboparty/roboparty_inference /tmp/roboparty_inference
git -C /tmp/roboparty_inference checkout 7a1178256d2b0f0e27427248076cfadc4db12f4d
node scripts/import-sim-motions.mjs /tmp/roboparty_inference
```

See `configs/beyondmimic.yaml`, `configs/getup.yaml`, and `configs/ufo.yaml`
in the corresponding upstream robot directories for inference contracts.
License: GPL-3.0, see POLICY-LICENSE.

## Locomotion policy

Source: https://github.com/Roboparty/roboparty_inference
Commit: ac448b12c382e3fab42c0e8467a6943987223c4b
File: robots/rpo/models/policy.onnx (unmodified)
Configuration: robots/rpo/configs/default.yaml (policy-config.yaml)
License: GPL-3.0, see POLICY-LICENSE.

Browser inference uses ONNX Runtime Web: 78 observations per frame, 10 history
frames (780 inputs), 23 actions, 50 Hz inference and 1 kHz physics. Joint order,
default pose, action scaling and PD gains follow upstream Sim2Sim and deployment
configuration. Velocity commands are bounded
to upstream deployment limits. This is a browser demonstration, not a hardware
controller. Simplified geometry can affect contact behavior relative to upstream.
