Skip to main content

Portfolio

Real jobs, rerun in physics

A real robot does a job; we rebuild the place as a physics twin and run the same job there, with the robot balancing on its own. Each case starts with the measured result.

Real → physics twin · 1 Delivered

Drinks into a fridge drawer, rebuilt and rerun in physics

Bottle in, drawer shut no fall in the 31 s run
Robot
Unitree G1 + five-finger hands
Source
One real recording
Delivered
Nov 2026

Problem

A real Unitree G1 puts a drink into a fridge drawer and closes it. Can the same job run in a physics twin, with the robot balancing on its own and nothing glued to its hands?

Approach

  • Counter, fridge drawer, bottle and cans rebuilt to scale from open asset libraries; every object collides with every other.
  • Joint angles, hands and torso taken from the real recording; arms planned around obstacles, legs left to the robot’s own balance controller.
  • The robot closes the drawer with its hand; the last part of the travel is a soft-close spring on the slide, as in real kitchen drawers.

Evidence

The bottle is released into the drawer and the drawer ends shut; the robot stays on its feet for the whole run.

Real → physics twin · 2 Delivered

Two drinks from a shelf into a basket, picked by friction

2 / 2 bottles grasped and placed in the basket
Robot
Unitree G1 + five-finger hands
Run
22 s, no fall
Delivered
Nov 2026

Problem

In the recording the robot leans 30° at the waist to reach the second bottle while its left hand carries the basket. An upright replay falls 13–26 cm short.

Approach

  • Torso lean taken from the recording and sent to the balance controller, so the hand reaches where the real one did.
  • Basket collision rebuilt as a hollow box and masses taken from the asset records (basket 0.4 kg, bottle 0.52 kg).
  • Bottles are free bodies held only by finger friction. Disclosed: the basket is fixed to the left hand, which stays closed on the handle in the recording.

Evidence

Both bottles in hand for about 4 s and released inside the basket. Right-wrist error against the recording 2.2 cm (median); deepest unintended robot–scene contact 2.0 mm.

Real → physics twin · 4 Delivered

A crumpled cloth picked up and put in a tray

100 % of the cloth inside the tray (target ≥ 50 %)
Robot
Unitree G1 + five-finger hands
Run
35 s, no fall
Delivered
Nov 2026

Problem

Cloth is soft: it bends, folds on itself and slips. The first attempt left only 43 % of it in the tray.

Approach

  • A 26 × 23 cm cleaning cloth from an open simulation-ready library, set to a realistic 30 g, simulated as a soft shell that can fold onto itself.
  • The cloth starts as a 7 cm crumpled pile; the right hand picks it up and hands it to the left hand, as in the recording.
  • Disclosed: while the recorded hand is closed, cloth points within 4.5 cm of the grip are held to it — friction alone cannot hold a cloth in simulation.

Evidence

The whole cloth ends inside the tray (checked frame by frame); the robot stays on its feet. Wrist error against the recording 1.1–1.5 cm (median).

Real footage: Unitree Robotics open datasets on Hugging Face (Apache-2.0).

Want this for your venue?

FOI rebuilds your hotel or café as a physics twin and trains the job there before it reaches your floor.

Accelerator
İTÜ ÇekirdekAcceleration 2026/4
Cloud program
Part of the Google for Startups Cloud Program
Immersion
Selected for Google for Startups Immersion EMEA, powered by Antler
Startup program
AWS Activate member
Investment program
Accepted into EIP 2026 · Global Entrepreneurship Festival, Cairo
Accelerator
İTÜ ÇekirdekAcceleration 2026/4
Cloud program
Part of the Google for Startups Cloud Program
Immersion
Selected for Google for Startups Immersion EMEA, powered by Antler
Startup program
AWS Activate member
Investment program
Accepted into EIP 2026 · Global Entrepreneurship Festival, Cairo