---
tags:
- microduck
- robotics
- reinforcement-learning
- onnx
library_name: microduck
pipeline_tag: robotics
---

# groove

In-place groove: rhythmic knee bounce with a small head sway (sim-trained, untested on hardware)

A **episodic** policy for the [microduck](https://github.com/pollen-robotics/microduck) (61-D observation, 14 actions, 50 Hz). Runs 10.0 s and returns itself to a standing pose.

## Run it on a robot

```bash
sudo robotctl policy add groove kramp/microduck-groove
robotctl robot do groove
```

The observation normalizer is baked into `policy.onnx`; feed raw observations.
`manifest.json` follows schema 2 of the microduck policy manifest (`docs/policy-manifest.md` in the daemon repo).

## Training

- **task_id**: `Mjlab-Dance-Flat-MicroDuck`
- **repo**: `pollen-robotics/microduck_rl`
- **branch**: `dance`
- **commit**: `2f3f4ea1d`
- **checkpoint**: `3999`

## What it actually does (measured in simulation)

20 s headless rollouts in CPU MuJoCo with the BAM XL330 actuator model (`scripts/render_policy.py`),
starting from the standing pose, all command slots at zero:

- **Never falls**: max trunk tilt 1.6°, drift 1 cm.
- **Knee bounce**: hip/knee/ankle bounce in sync on both legs, at 87% of the reference amplitude,
  about 2.1 Hz (≈126 BPM rather than the 100 BPM target).
- **Head**: a small head-yaw sway at 30% of the reference amplitude. The planned head nod, head tilt
  and hip sway were **not** learned.

`replay.mp4` shows it.

## How it was trained

- Task `Mjlab-Dance-Flat-MicroDuck` (mjlab, PPO), built on the velocity recipe: domain randomization,
  pushes, and BAM actuators, with the twist command always zero.
- The actor gets no clock (the standard 61-D contract), so rhythm has to come from state feedback.
  The reward runs a phase-locked reference: it advances at the target tempo and may slip only a little,
  so standing still never pays.
- Reference-state initialization: 70% of episodes start mid-dance.
- 4000 iterations × 4096 envs on one L4 (HF Jobs, about 2 h 50 min).
- The training code is a local fork of `pollen-robotics/microduck_rl` (branch `dance`, not published),
  so the commit hash below does not exist upstream.

## Caveats

- **Sim-only. It has never run on a real robot.** Test it held or on a soft surface first.
- The policy grooves indefinitely. When `duration_s` (10 s) ends, the daemon hands back to the stand
  policy mid-bounce rather than from a settled pose.