The machine above is a study. What follows is the work behind it: how it will think, what is being researched, the order in which it will be built, and how that work is funded.
Study No. 09
The Mind
Morph is designed to build on open vision-language-action models rather than a closed controller of its own. A camera image and a plain sentence go in, and joint commands come out.
Fig. 9.1 — Perception, reasoning, action. Drawn as a design target; Morph's own integration is unbuilt.
Reference I
OpenVLA
An open-source vision-language-action model with 7 billion parameters, released by researchers from Stanford, UC Berkeley, Google DeepMind and the Toyota Research Institute. Its visual encoder fuses DINOv2 and SigLIP features. These are projected into a Llama 2 language model that predicts robot actions as tokens. It was trained on about 970,000 robot episodes from the Open X-Embodiment dataset, and its weights and training code are public.
Reference II
π0 (pi-zero)
A vision-language-action model from Physical Intelligence, built on the open PaliGemma vision-language model. It adds a separate, smaller action expert trained with flow matching, which produces smooth, continuous chunks of motion rather than discrete tokens. Code and weights are open-sourced in the openpi repository.
Approach
Fine-tune an open VLA on Morph's own body and hand · design target
Inputs
Head cameras, joint state, spoken or written instruction · design target
Outputs
Joint and tendon commands for hand, arms and legs · design target
Below the model
Classical balance and safety controllers remain in charge of the body · design target
Five open lines of inquiry. None is solved here. Each is listed with the question it asks.
i.
Dexterous manipulation
Can a tendon-driven hand, with its motors in the forearm, give fine and gentle control of fingertip force? The aim is a hand that holds an egg and a hammer with the same mechanism.
Open research question
ii.
Whole-body control & balance
Keeping a standing body upright while the arms reach, lift and push. Contact forces at the feet are read continuously and the posture is corrected below the level of conscious planning.
Design target
iii.
Sim-to-real training
Most practice happens in simulation, where falling is free. The question is how much of that practice survives contact with real friction, cable stretch and sensor noise.
Open research question
iv.
Learning from demonstration
Household tasks taught by showing rather than programming: folding, clearing, opening and carrying. Recorded demonstrations become fine-tuning data for the model in Study No. 09.
Design target
v.
Safety & compliance
Backdrivable joints, torque limits and an always-available stop, so the machine yields when it touches a person. Safety is treated as a property of the hardware first and of software second.
Design target
Pl. 10.1–10.2 — Studio studies, hand and head. Concept imagery, not a built prototype.
The research is funded in the open. Creator fees from trading in the project's token go to the project's development wallet and fund the milestones above one at a time, and each result is published.
Fig. 12.1 — Flow of funds, in principle.
Principle
Funded in order
Fees pay for one milestone at a time. The next is not begun until the last is published.
Principle
Published openly
Designs, findings and failures from each milestone are released for anyone to study.
Morph is an experimental research project. Nothing here is financial advice or a promise of any return.
Pl. 12.1–12.2 — Front and back studies. Concept imagery.
