Totallytubular.

Walking
before soaring.

We're landing a modified Berkeley Humanoid Lite first. Then we pivot to a bespoke platform.

  1. NowTK1Modified Berkeley Humanoid Lite
  2. NextTK2Bespoke platform. Significantly more agile, more robust, and acrobatic.

TK1

A modified Berkeley Humanoid Lite.

Height
1.04m
Mass
19.5kg
Actuators
22
Per hand
0.5kg

Legs

Joints
Six per leg
Walking
Learned policy

Arms

Joints
Five per arm
Hands
SO-101 grippers

Head

Depth
RealSense D435i
Eyes
Colour stereo pair

Torso

Compute
Intel N100
Stop
Hardware e-stop

Any model can drive it.

tk1 · mcpt+00.0 s
  1. youBring me the red bottle on the table.

  2. stand()

    Done: standing

  3. point_head(pan_degrees=18, tilt_degrees=22)

    Done: head at pan 18, tilt 22 degrees

  4. look(camera="depth")

    image/jpeg 640 × 480

  5. The red bottle is at (412, 233), left of the cup.

  6. locate(u=412, v=233)

    Done: x 2.71 m ahead, y 0.86 m left, z 0.12 m above the hips

  7. walk_to(forward=2.45, left=0.74, heading_degrees=18)

    Done: walked 2.56 m by dead reckoning

  8. look(camera="depth")

    image/jpeg 640 × 480

  9. locate(u=301, v=338)

    Done: x 0.26 m ahead, y 0.13 m left, z 0.09 m above the hips

  10. pick(x=0.26, y=0.13, z=0.09, width=0.06)

    Done: holding it in the left hand

  11. turn(degrees=90)

    Done: walked 0.04 m by dead reckoning

  12. turn(degrees=90)

    Done: walked 0.03 m by dead reckoning

  13. walk_to(forward=2.6)

    Done: walked 2.61 m by dead reckoning

  14. say(text="Here you go.")

    Done: spoken

KineBench

A novel kinematics benchmark for evaluating the performance of frontier foundation models and specialized robotics models in physical AI applications.

  1. 01

    A task in words

    Each episode is a scene and an instruction, drawn from a seed.

  2. 02

    The model gets MCP

    It drives the robot with the same tools that run it in the real world.

  3. 03

    The simulation scores it

    From ground truth, never from what the model says it did.

pick-named

“Pick up the yellow block and hold it.”

ScoredIs the yellow block held up?

Research

Creating extensible, highly applicable, backwards-compatible robotic platforms to push the frontier of physical AI.

  1. Hands

    Sophisticated hands, fingers and other appendages.

  2. Quantum

    Quantum data links, and quantum computing in robotics.

  3. Systems

    Interdisciplinary research across hardware, software and systems.

Totally tubular.