---
tags:
- reinforcement-learning
- locomotion
- bipedal
- microduck
- pollen-robotics
---

# Microduck RL — 2.0 m/s running push

Fork of [pollen-robotics/microduck_rl](https://github.com/pollen-robotics/microduck_rl) training
Pollin Robotics **Microduck** (25 cm open-source biped, XL330 servos) to run at a
**measured mean 2.0 m/s** forward speed via PPO (rsl_rl + MuJoCo Warp / BAM).

## Honest status (2026-09-12, TRAINING STOPPED)

- Recipe: exact DuckEMW running config, **only deviation = speed cap raised to 2.5**
  (`MICRODUCK_RUNNING_SPEED_CAP=2.5`, `MICRODUCK_RUNNING_TARGET_MAX_SPEED=2.5`).
  No flight reward, no Su-style rewards, no achievement gate — reverted to the proven base.
- **Max measured mean forward speed (world-displacement eval): 1.621 m/s @ cmd 2.0**
  (p10 1.387, p90 1.828, max 1.923, ~2 resets) at checkpoint `model_76500.pt`
  (iter 76506). This is the best result of the run.
- Earlier best was 1.651 m/s @ cmd 2.0 (model_58000, cap 2.0). The cap-2.5 run reached
  1.621 mean at the final checkpoint; fluctuated 1.52–1.65 across the climb — did not pass 1.7.
- Goal (mean 2.0 @ cmd 2.0) **NOT reached**. Training halted at user request.

## How speed is measured (no lies)

`scripts/eval_sprint_speed.py` measures true world-frame displacement
(`root_link_pos_w` delta / time) across 64 envs, reporting mean / p10 / p90 / max / resets
per command. Reward/curriculum metrics are NEVER reported as speed.

## Full honest eval — model_76500.pt

| cmd (m/s) | mean  | p10   | p90   | max   | resets |
|-----------|-------|-------|-------|-------|--------|
| 0.4       | 0.078 | 0.000 | 0.174 | 0.377 | 12     |
| 0.8       | 1.211 | 0.636 | 1.565 | 1.615 | 15     |
| 1.2       | 1.512 | 1.187 | 1.775 | 1.858 | 4      |
| 1.6       | 1.575 | 1.323 | 1.820 | 1.899 | 6      |
| 2.0       | 1.621 | 1.387 | 1.828 | 1.923 | 2      |

## Run it

```bash
export MICRODUCK_RUNNING_SPEED_CAP=2.5
export MICRODUCK_RUNNING_TARGET_MAX_SPEED=2.5
.venv/bin/train Mjlab-Running-Flat-MicroDuck \
  --env.scene.num-envs 4096 --agent.max-iterations 80000
```

## Files

- `src/mjlab_microduck/tasks/microduck_running_env_cfg.py` — running task cfg + knobs
- `src/mjlab_microduck/tasks/mdp.py` — rewards + speed curriculum
- `scripts/eval_sprint_speed.py` — lie-proof world-displacement eval
- `checkpoints/model_76500.pt` — final checkpoint (best: 1.621 m/s @ cmd 2.0)
- `checkpoints/model_59000.pt` — earlier cap-2.0 checkpoint (1.65 m/s @ cmd 2.0)