Physical Autocomplete for Construction Robots

An agent harness for the physical world.

PACR orchestrates planning, model execution and failure recovery for humanoid robots. It turns a one-minute demo into an unattended shift.

Leipzig · Europe
01 / The problem

Errors are certain. Giving up isn’t.

No robot gets through a night without mistakes. What matters is what happens next.

Single modelno one catches the error
PACRerror caught, fixed, verified
step doneerrorerror recovered
Why it matters: at 99 % per step, only 37 % of 100-step runs finish.
02 / The harness

Four kinds of AI. Each doing what it does best.

No single model can plan a night, decide in a split second and re-grip a slipping bag. PACR gives each moment to the model built for it.

← slower · sees the whole missionfaster · controls one motion →
Seconds

LLM

Plans the mission

Turns one sentence into a task plan. Replans when the world changes.

0.1 to 0.5 s

System One

Decides what’s next

Picks the right skill for this moment, or hands back to the planner.

Tens of ms

VLA

Handles the general case

A vision-language-action model for tasks no specialist covers yet.

Milliseconds

RL policies

Master the specifics

Trained specialists: walk on gravel, grip, carry, re-grasp.

Safety supervisorIndependent of all models. Stops everything, any time. The LLM never touches a motor.
03 / Failure recovery

Something fails. The shift goes on.

System One decides in under half a second: switch to a specialist policy, replan with the LLM or ask a person one question.

Error detected

Bag slips from the grip

VLA · carrying bag 23

System One< 0.5 s
Switch policymilliseconds
VLA → trained RL specialist
Replanseconds
LLM writes a new plan
Ask the operatorrare
One question, no takeover

Most errors are fixed in milliseconds. Only the hard ones reach the LLM, and only the rare ones reach a person.

Illustrative timings
Real footage
On the robot

The box doesn’t go in on the first try.

The robot notices, pulls back, re-grips and still gets it onto the shelf. That is the whole idea, at the smallest scale.

04 / The goal
10×

more autonomous runtime and task completion than the best single model.The models already exist. What’s missing is the system.

Best single modelsame hardware
1×
PACR harnesstarget
10×
8 h
night shift as the benchmark
1
sentence to start the mission
0
teleoperation
First sites
Rock sites with Atelier Thilo Krause: slopes no wheel loader reaches, cement bags carried by hand. Stores at night with Niklaus Baugeräte: inventory and restocking while closed.
05 / Team

Built in Leipzig.

Robotics engineers who run autonomous humanoids at RoboCup, and builders who know what a site actually needs.

HTWK Robots

HTWK Leipzig

Two-time RoboCup world champions. Silver in humanoid soccer at the World Humanoid Robot Games in Beijing, 2025 and 2026. Since 2009 the team has developed the AI that lets robots act on their own in real time: RL locomotion, perception, VLAs, strategy, planning and multi-robot coordination.

The harness, the models, the robot software.

Havenburg Media

Leipzig

Builds the rock landscapes and themed worlds that Atelier Thilo Krause designs for zoos and parks across Europe. Steel, sprayed concrete and handwork, no two rocks alike. This is where today’s robots give up.

The sites, the work, the market.
HTWKT
Thomas
Architecture & product
HTWKE
Erik
Harness workflow
HTWKF
Felix
Harness & data platform
HAVENBURGA
Arne
Business development
Partners & first customers

Atelier Thilo Krause

Partner · customer 1

The name behind rock worlds across Europe since 2001: the Lewa Savanna at Zoo Zürich, the polar bears at Schönbrunn, Zoo Leipzig, Europa-Park.

Use case: carrying material on rock sites no machine reaches.

Niklaus Baugeräte

Customer 2

Construction equipment sales and rental for over 50 years, with eight branches across Baden-Württemberg.

Use case: inventory and restocking at night, while the store is closed.
Physical autocomplete & recovery

Prompt the change you want to see in the world.

›
Tab ⇥ accept