Morph · Atelier

Atelier

The lab itself: what we intend to build, in what order, how we will know if it worked, and how the work is paid for. Everything on this page about Morph is a design target — a plan, not a result.

The programme · principles
Milestones M1–M6 · experiments and exit criteria
Lab log · what has actually happened
Open questions
Tokenized R&D · how the work is funded
Reading chains
Glossary

The programme

Build on open work

We start from published, reproducible designs — the LEAP Hand, LeRobot, openpi, Isaac Lab — rather than inventing what already exists.

Upper body first

Most household value is in hands and arms. Legs come last (M5), following the reasoning in Study 04.

Publish everything

CAD, code, data and failures are released at the end of each milestone, so the work can be checked.

One milestone at a time

The next milestone is not begun until the last is published.

Milestones

Each milestone is an experiment with a question, a method built on cited prior work, and an exit criterion. No costs or dates are given: we have not sourced numbers we can stand behind, and we would rather print none than invent them.

Hand prototype design target

Question: can a LEAP-style direct-drive hand be modified to move its heaviest actuators into a forearm without losing the controllability that makes LEAP useful for learning?

Build a stock LEAP Hand from its published design as a baseline.
Build a variant with tendon-driven flexion for two fingers (tendon drive, as in the Shadow Hand).
Measure fingertip force, repeatability and finger mass on a bench for both.
Add a fingertip DIGIT or AnySkin sensor to one finger (Study 02).

Exit: published CAD, bill of materials and the bench comparison, whichever design wins.

Forearm + wrist design target

Question: what wrist drive gives enough torque while staying backdrivable?

Compare a quasi-direct-drive wrist (MIT Cheetah style) to a cycloidal one.
Test contact force against ISO/TS 15066 limits for hand and arm.

Exit: a 3-DoF wrist carrying the M1 hand, with published test data.

Upper body + teleop rig design target

Question: can we collect useful household demonstrations cheaply?

Two arms on a fixed stand, teleoperated with an ALOHA-style leader/follower setup.
Record in LeRobot format; add handheld UMI-style collection for robot-free data.
Target tasks drawn from BEHAVIOR-1K: folding, wiping, loading a dishwasher.

Exit: an openly licensed dataset of household episodes, collected with consent and scrubbed of personal information.

VLA integration design target

Question: how much of our own data does an open VLA need before it handles our tasks?

Fine-tune π0 / π0.5 via openpi on M3 data; compare against OpenVLA-OFT and an ACT baseline.
Sweep demonstration count per task; report success curves.
Sanity-check the pipeline on LIBERO first so results are comparable with published numbers.

Exit: released fine-tuned weights, eval scripts and a report including failures.

Full humanoid design target

Question: can learned locomotion carry the upper body safely?

Train walking and standing balance in Isaac Lab or MJX with domain randomization, following Rudin et al. and Radosavovic et al.
Evaluate on an open platform such as Unitree G1 before custom legs.

Exit: sim-to-real report and released policy.

Home pilot design target

Question: what goes wrong in a real home?

Supervised trials only: a person present and a physical stop at all times.
Score semantic safety with ASIMOV-style scenarios and follow ISO 13482 as the design reference.

Exit: a public incident and failure log alongside success rates.

III

Lab log

Only things that have actually happened are recorded here.

CompletedStudies 01–07 and the Catalogue published: a sourced review of hands, touch, actuation, locomotion, VLA models, data and safety.
CompletedResearch programme M1–M6 defined as design targets.
NextBegin M1: source parts for a baseline LEAP Hand build.

Open questions

How long do vision-based and magnetic tactile skins last under household wear — water, detergent, heat? We found no long-term published study. (Study 02)
Is the finger-mass saving of tendons worth their friction and maintenance in a home robot? (Study 01)
How many household demonstrations per task does a fine-tuned π0.5 or OpenVLA-OFT need on a new embodiment? (M4)
Can robot-free collection (UMI) replace most teleoperation for dexterous hands, not just grippers?
What is a fair, reproducible home benchmark — and how do we report failures honestly? (Study 07)

Tokenized R&D

The research is funded in the open. Creator fees from trading in the project's token go to the project's development wallet and pay for the milestones above, one at a time.

Fig. V.1 — Flow of funds, in principle.

Funded in order

One milestone at a time. The next begins after the last is published.

Published openly

Designs, data, weights and failures from each milestone are released.

No promises

Research can fail. A milestone that fails is still published as a result.

Morph is an experimental research project. Nothing here is financial advice or a promise of any return.

Reading chains

Follow a chain from first paper to frontier. Each link opens its catalogue entry.

VII