Two robots. One idea: turn text into measured contact, not just an animation.
4× preview · Original-speed dual showcase · G1 full video · G1 recorded states + video
WriteUrRobo brings DexMate Vega-1P and Unitree G1 + Inspire hands into a reproducible ASCII writing showcase. Both use fixed-base, gravity-compensated simulation with an attached compliant-nib pen. Ink is reconstructed from actual nib–paper contacts, not painted from the intended letters.
| DexMate Vega-1P | G1 + Inspire | |
|---|---|---|
| Controller | Tip-position and normal-force feedback | Position servos + bounded Cartesian integral correction |
| Pen | Rigid attachment, compliant nib | Rigid fixture, compliant nib |
| Model | Dexmate official assets / calibrated station | Pinned Unitree revision / DFQ Inspire hands |
| Reproduce | DexMate instructions | G1 method and commands |
Research scope, not a hardware claim: neither demo proves autonomous grasping, safe pen pressure, collision-free execution or unsupported standing balance. The controllers and stations differ; this is a showcase, not a controlled robot benchmark. Original DexMate commands and releases remain available.
The complete G1 Hello, Lab. run lasts 68.87 s: mean pen-down XY error 0.955 mm, P95 1.673 mm, contact fraction 95.08%, peak recorded nib force 0.396 N. The tracking/contact checks pass for this example only. Contact during 440 nominal pen-up samples is retained in the ink. Enabled-contact overlap remains, including a 5.20 mm thumb/fixture mesh contact; these are not collision-clearance or hardware-safety results. Metrics · Checks and scope.
git clone https://github.com/MrAmazing-h/WriteUrRobo.git
cd WriteUrRobo
python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
g1-fetch-writing-model --output external/g1
g1-contact-write --model-dir external/g1 --text "Hello, Lab." --output outputs/g1_hello
MUJOCO_GL=egl g1-render-writing --output outputs/g1_helloThe distribution/import name stays dexmate-writing / dexmate_writing for compatibility. G1 adds independent commands without changing the original controller. See station assumptions, data schema and limitations.
4× preview · Watch / download the full 77.58-second video at original speed
The original release animated a fixed Chinese phrase. The new pipeline accepts printable ASCII letters, numbers, punctuation and explicit line breaks, wraps and scales the text, generates a smooth joint reference, and executes it with MuJoCo physics. Pen-tip position and normal-force feedback correct the reference while the left hand maintains contact with the table. Ink comes from recorded nib–paper contacts, including unintended pen-up contacts, rather than painting the planned font paths.
Scope: the pen remains rigidly attached to the right hand with a passive compliant nib. The base is fixed and the robot links use gravity compensation. This is contact-writing simulation, not friction-only grasping, autonomous pickup-to-writing, tactile grasp adaptation, RL, or hardware validation. Inherited collision exclusions remain, and impact-force peaks are unresolved.
| Example | Full video | Result |
|---|---|---|
Single-line Yuxiong Lab |
45.29 s, original speed | Overhead image |
Two-line DexMate 2026! / Hello, Lab. |
77.58 s, original speed | Overhead image |
| Original Chinese phrase, kinematic | v0.1.0 video | Legacy method and commands |
The single-line example uses the earlier calibrated letter paths; the configurable Hershey layout was evaluated separately on the two-line example. The photos show archived physical-rollout states. They do not establish autonomous grasping.
Linux and Python 3.11 are the locally tested setup. Use a new output directory for each run. Rendering needs a working OpenGL backend; FFmpeg is supplied by imageio-ffmpeg.
git clone https://github.com/MrAmazing-h/WriteUrRobo.git
cd WriteUrRobo
python -m venv .venv
source .venv/bin/activate
pip install -e '.[dev]'
# Layout only: no robot model, simulator rollout, or GPU required.
dexmate-contact-write --text "Hello 123!" --output outputs/layout --layout-only
# Fetch the calibrated station (about 14 MB, SHA-256 checked).
dexmate-fetch-contact-station --output external
printf 'DexMate 2026!\nHello, Lab.\n' > text.txt
MUJOCO_GL=egl dexmate-contact-write --text-file text.txt --output outputs/contact--station-dir selects another copy of the same calibrated station; it is not an adapter for arbitrary robots. --max-cap-mm 25 reproduces the recorded two-line layout setting. --no-render runs planning, physics and evaluation without video; MUJOCO_GL=osmesa can be used where OSMesa is installed. Each stage writes a log before reporting completion. The model is shared through a relative symlink, so keep the station directory in place.
All 95 printable ASCII characters are represented by Hershey futural; spaces have no strokes. Chinese text is rejected by this interface. Word wrapping, long-word splitting and scale fitting use a calibrated 280 × 70 mm area, with a minimum capital height of 10 mm. Text that does not fit is rejected. Layout support for a character does not certify its physical execution.
| Measurement | Single line | Two lines |
|---|---|---|
| Mean pen-down XY tracking error | 0.264 mm | 0.257 mm |
| P95 pen-down XY tracking error | 0.505 mm | 0.503 mm |
| Stroke coverage within 1 mm of measured ink | 100.00% | 100.00% |
| Pen-down contact coverage at 100 Hz | 99.699% | 99.939% |
| Maximum enabled-contact penetration, all 2 kHz substeps | 0.603 mm | 0.725 mm |
| Peak individual recorded nib contact force | 12.826 N | 21.270 N |
These are two specific simulation demonstrations, not a success rate over random texts. The force peaks are unresolved impact transients; passing geometric acceptance does not establish safe pen pressure. Coverage is the nearest same-stroke physical ink distance from a densely sampled reference path; it differs from time-aligned tracking error. The old kinematic IK error and the new physical-rollout error are not an apples-to-apples improvement comparison.
See full two-line metrics, acceptance, all-substep collision audit, two-example CSV, and method, feedback equations and reproduction scope.
layout.json/layout_preview.png: planned layout only.reference.npz: 100 Hz joint/pen-tip reference, stroke IDs and pen-down mask.rollout.npz: actual simulated joint states, contact forces and ink samples, plus the reference index (which can pause while contact is established).metrics.json,substep_collision_audit.json,acceptance.json: evaluated geometric criteria and limitations.videos/text_contact_writing.mp4/videos/final_overhead.jpg: physical-state replay with measured-contact ink.
Release v0.2.0 includes videos, the calibrated station, and a compact recorded reference/rollout bundle. These NPZ files are simulation records, not certified deployment commands. The v0.1.0 Chinese example and dexmate-writing CLI remain available unchanged.
ruff check .
ruff format --check .
pytest -q
python -m buildCode: MIT. Robot assets: Dexmate Inc., Apache 2.0. ASCII layout: Hershey-Fonts 2.1.0 / Hershey vector fonts. Legacy Chinese stroke data: Hanzi Writer Data 2.0.1, with its bundled Arphic license. See third-party notices, contributing, and citation.
Related project: DanceUrRobo — video-to-robot dance demonstrations.