SIMULATION
Supabotics · open robotics

Robots that walk.
Robots that fold.

Two robots, built in the open. Every result is dated, sourced and labeled.

Read the log
SIMULATION
Project 01 · Simulation stage

RoboDuck

A small biped built on Pollen Robotics' Microduck and Open Duck Mini, trained with reinforcement learning in MuJoCo.

View project
SIMULATION
Project 02 · Offline, no hardware yet

LaundryBot

Two SO-101 arms, fine-tuning open models on public towel-folding data. First target: fold a hand towel twice.

View project
Results

Numbers, with sources.

RoboDuckSIMULATION
101.5%
Velocity tracking at 0.1 m/s
± 2.1 · 3 seeds × 10 runs · 0 falls
microduck/results.tsv:9 →
RoboDuckSIMULATION
0.259 m/s
Measured speed at a 0.3 m/s command
Stage B · nominal randomization · 10 trials · 0 falls
microduck/results.tsv:5 →
LaundryBotSIMULATION
16/20
Scripted towel two-folds passed
Newton VBD cloth · scripted motions, not a learned policy · Wilson 95% 0.58–0.92
laundrybot/rounds.md:41 →
LaundryBotOFFLINE
0.1321
10-step action error, single dataset
Offline, recorded data · pooled 4 datasets: 0.1442 · hold-position: 0.1744 · lower is better
laundrybot/rounds.md:104 →
Log

Latest updates

2026-10-23RoboDuck walks in simulationRoboDuckSIMULATION
2026-10-23LaundryBot: pooling four towel datasets didn't help (yet)LaundryBotoffline

All updates