AI training for humanoid robots

Doesn't do
what you want yet?

We teach humanoid robots new skills: recording demonstrations, teleoperating them by hand, and fine-tuning open AI models like GR00T and OpenVLA on top. AI training and development, for Unitree owners and developers.

LEROBOT · UNITREE G1 OFFICIALLY SUPPORTED GR00T · OPENVLA · π0 FINE-TUNING TESLA OPTIMUS — WE ARE WATCHING IT CLOSELY
Two chapters

What we do

Chapter 1 — Demonstration

Demonstration

We record how a task should be done. No expensive rig required to start: a webcam and open-source hand-tracking already produces usable demonstrations. We scale up to VR or a puppeteering rig if your project needs it.

  • Webcam-based teleoperation, no headset needed
  • Affordable practice hardware (SO-101, ~€150–300)
  • VR teleoperation when you already have the gear
  • No robot yet? We start from public datasets
We work with: Unitree G1, R1, Go2 · Booster T1 — and we're preparing for Tesla Optimus
Chapter 2 — Training

AI development

We fine-tune open foundation models — GR00T, OpenVLA, π0 — on your demonstrations using Hugging Face's LeRobot, or train focused policies with reinforcement learning in simulation. Real skills, not scripted motions.

  • Fine-tuning on GR00T, OpenVLA, π0 / π0.5
  • LeRobot-based training pipelines
  • Public datasets (Open X-Embodiment) or your own
  • Reinforcement learning & sim-to-real (Isaac Sim, MuJoCo)
Booked per block of hours — tell us what skill you want to teach.
How robots learn

Three ways to teach a robot

01

Show it

The simplest method: physically guide the robot's arm through the motion (kinesthetic teaching), or let it copy a demonstration. Cheap and direct, but it only scales to simple, repetitive tasks.

02

Teleoperate it

A human drives the robot remotely. The accessible way in: open-source hand-tracking turns an ordinary webcam into a controller — no headset needed. From there you can scale up to a low-cost arm like the SO-101 (€150–300), a VR headset (Unitree's own xr_teleoperate supports Vision Pro, PICO and Quest), or a puppeteering rig like GELLO. Every recorded run becomes training data.

03

Fine-tune a foundation model

The modern approach: instead of training from scratch, start from a model already pretrained on hundreds of thousands of robot demonstrations — GR00T (NVIDIA), OpenVLA, or π0 / π0.5 (Physical Intelligence) — and fine-tune it on a few hundred of your own, using Hugging Face's LeRobot toolkit. Days, not months.

Have a skill you want your robot to learn?

Contact