Dhanush RadhakrishnanFounder & CEO

Our vision for a musculoskeletal clone comes from a deep desire to make an android that can do anything a human can do. I grew up watching Boston Dynamics and, after GPT arrived, it became clear deep learning works and can scale. That made a high degree of freedom android feel tractable.

The mission is complete androids, but the hand is the hardest part. It has the most degrees of freedom and many to one and one to many muscle joint links. We proved a durable human level hand in 18 months, then used the same design systems to move fast to the full body.

If artificial muscles match human skeletal muscle in displacement, force, and speed, then an anatomically accurate android will match human range of motion, strength, and joint speed. That lets us reference anatomy textbooks one to one instead of engineering every joint from scratch.

The bar is human parity across strength, speed, dexterity, and a soft, fleshy body at a cost we can produce. Copying nature reduces the number of problems. We focus on making muscles and polymer analogs for soft tissues, then attach them like in the body.

Not now. First we want human parity. That alone is very hard. Once we have a human level body and a powerful brain, the android can do anything a human can do, and later probably more.

We keep tech in house and continue to verticalize. We avoid retrofitting with industrial arms that would distract from the vision and dilute the uniqueness of our artificial muscle product.

The clone as the ideal human companion. It fits into daily life like a perfect puzzle piece, helping at home and at work, from simple tasks like laying out clothes to acting as an assistant on the job.

It is efficiency times total energy. First prototype: about one hour of full speed running before the battery drops. In homes, where it manipulates more and sprints less, that should translate to several hours. We expect to push the number higher over time.

Both verbal and non verbal. Early on, users will demonstrate tasks with their own hands. With action labels and language labels, a vision language action model lets you tell it what to do. Over time it anticipates needs.

Motors force trade offs and complex packaging inside a rigid shell. With muscles close to human muscle, we have an existence proof for human level specs. Manufacturing is also simpler for us: we auto produce muscle fiber by the kilometer and attach to bones at the right lengths.

It is a fluidic McKibben muscle. For prototyping we sometimes use air. For the untethered product we use water with a compact hydraulic pump inside the torso. The pump is like a hydraulic heart.

Train in a physics simulator with procedurally generated terrains and domain randomization, then transfer to the real robot. Our bottleneck is a faithful clone of our soft, muscle driven robot in a GPU simulator.

Today we use Nvidia Jetson in the skull. If we can fit two, one can plan and one can do motion control. Over time we may add ASICs across the body like neurons near key organs.

Joint positions and torques, muscle length and force, plus vision and audio. We delayed full tactile skin for now because position control is the dominant paradigm, but whole hand high resolution tactile maps will be very useful later.

Collect a large and diverse set of bimanual manipulation data with language labels. Train a robot foundation model so the robot generalizes to new tasks after a few user demonstrations.

Skin for appearance so it is not frightening, and a first foundation model that learns new manipulation tasks from a handful of demos. Users will show it, then it should perform the task smoothly.

Untethered bipedal walking. It needs compact valves, batteries, the hydraulic heart, accumulators, reservoir, tubes, and cables all inside rib cage, abdomen, and pelvis, plus trained locomotion. Hitting this proves a true synthetic human that can move on its own, which unlocks real world value at scale.