Open source · Raspberry Pi · 59 tests green

Build your own TARS.

A walking, talking robot from Interstellar — with a brain, a sense of humor, and a gait it learned by itself.

Get the code

Drag the robot to orbit · click a part to inspect it

EXPLODE
·
€280entry cost, printer owned
21plugin skills
18,500gaits tried per hour, in sim
€0/mofully-local mode
01

Systems

One Python process on a Raspberry Pi 5. Every layer degrades gracefully — from a full cloud AI stack down to 100% offline and free.

/01

Hands-free voice

Offline wake word, a spoken "Yes?", follow-up questions without repeating the name, replies streamed sentence-by-sentence while the LLM is still thinking.

/02

Plugin skills

Drop a decorated Python function in a folder, it becomes an LLM tool: move, look, remember, timers, music, Home Assistant, image generation.

/03

Personality dials

Humor, honesty and sarcasm from 0 to 100 — adjustable by asking. Character cards switch TARS into CASE or KIPP.

/04

Balance first

No arms, by design. An IMU fall watchdog relaxes the servos when TARS tips over, and a gyro stability tax makes smoother gaits win during training.

/05

Memory & knowledge

Semantic long-term memory plus a structured knowledge graph, injected into conversation only when relevant.

/06

Anywhere dashboard

Streaming chat, browser voice from any phone on the LAN (HTTPS built in), movement pad, live learning curve — and a movie-style onboard readout at /display.

/07

Cloud or fully local

OpenAI + ElevenLabs for the best experience, or Ollama + Piper + Vosk for a TARS that never phones home.

/08

Builder tooling

tars --doctor self-tests 12 subsystems with fix hints, --benchmark times the voice pipeline, and everything runs in simulation on a laptop.

02

The build

Raspberry Pi 5, a PCA9685, three to six servos, ~1.5 kg of PETG and a weekend of printing. Bilingual docs (EN/IT) cover everything — including which Bambu Lab printer to buy and how to get the brushed-metal movie finish. Chassis STLs by the TARS-AI Community.

03

It learns to walk

No simulation hand-waving — the reward is real centimeters covered. An evolution strategy mutates the gait timing; the tape measure, the camera or the IMU judges the result. A MuJoCo pre-filter (the same gait code, driving physics instead of servos) screens thousands of candidates overnight, and a Spearman honesty check tells you exactly how much to trust it on your floor.

# overnight, on a PC — 6 randomized worlds per candidate
$ python -m tars.learn --reward mujoco --iterations 300 --dr 6

# how honest is the sim about YOUR build?
$ python -m tars.learn --correlate
  Spearman rank correlation sim vs real: rho = +0.71
  Verdict: strong — the sim is a useful pre-filter for this build

# then verify the winner on the real floor
$ python -m tars.learn --reward camera --steps 3
04

Origins

TarsGPT is based on Charlie Diaz's original TARS replica — the idea that started it all — rebuilt as an original codebase and extended with everything on this page: the AI runtime, the skills system, gait learning with verifiable rewards, the balance-first redesign and the bilingual documentation. The 3D-printable chassis is by the TARS-AI Community (CC BY-NC 4.0). TARS is a character from Christopher Nolan's Interstellar; this is a non-commercial fan project, unaffiliated with the film.