---
title: "To control an excavator, MIT built a miniature one"
locale: "en"
url: "https://irz.fr/en/articles/excavator-world-space-interface-en"
markdown_url: "https://irz.fr/en/articles/excavator-world-space-interface-en.md"
category: "tech"
tags: ["MIT", "robotics", "interface", "excavator", "construction", "teleoperation", "HCI"]
published_at: "2026-08-22T09:18:00.000Z"
author: "Camille Morel"
translation: "https://irz.fr/fr/articles/excavator-world-space-interface-fr.md"
---

# To control an excavator, MIT built a miniature one

Two joysticks force an operator to translate intent into four machine joints. MIT's World-Space Interface copies the excavator geometry so software can perform that translation instead.

MIT's answer to a difficult excavator control problem is sitting on a desk and looks suspiciously like a toy excavator.

The **World-Space Interface**, or WSI, has a rotating base, two linked arms and a hand control for the bucket. An operator grabs the little mechanism and moves it roughly as the full-size machine should move. Behind the scenes, encoders and software turn that gesture into commands for the simulated excavator.[1](https://news.mit.edu/2026/mit-engineers-design-better-controller-operating-construction-diggers-0820)[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488)

That sounds almost too obvious, which is part of the appeal. Conventional excavators already have a visible mechanical logic, yet their controls ask the operator to learn a separate language of two joysticks. The MIT and Sumitomo team has spent several years asking whether that translation belongs in the operator's head at all.

## Four joints

A hydraulic excavator has four primary degrees of freedom in the study: cab rotation, boom, stick and bucket.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488) The joysticks command those joints, not the bucket tip directly.

Imagine the job in front of the cab: the bucket needs to move diagonally towards a pile, curl, scoop and then swing away, a path that is easy to see from the seat but awkward to express through two joysticks because several joint velocities have to cooperate.

Researchers call the learned conversion an **internal model**. The map connects a goal out in the world with the motor commands needed to reach it.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488) Experienced operators have practised the conversion until it disappears from conscious thought, whereas a newcomer receives the whole problem at once.

That distinction matters. Skill at excavation and skill at speaking the joystick dialect overlap, but they are not identical things.

> **Two routes to the bucket**
> Diagram comparing the mental mapping required by dual joysticks with the World-Space Interface.
> - Two joysticks
> - Mental map: cab + boom + stick + bucket
> - Bucket
> - Miniature WSI arm
> The WSI does not remove kinematics; it moves the conversion from the operator's brain into the interface and software.

The WSI shifts that burden. Its linkage is a scaled mechanical echo of the excavator itself, so the operator can think about where the bucket should go while the system handles more of the joint-space translation.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488)

Hence the slightly academic name. **World space** is the space of the pile, trench and rock outside the cab; **joint space** is the set of individual machine articulations that must cooperate to reach them.

## Small machine

The hardware has pleasingly ordinary parts. The published design uses a vertical rotary joint for cab rotation, a parallel linkage for boom and stick, and a bicycle brake lever to control the bucket.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488)

The 2025 precursor describes two 240 mm links, a 46 mm third link, springs that offset the mechanism and the expected weight of an operator's arm, plus rotary and linear encoders read at **300 Hz**.[3](https://construction-robots.github.io/papers/72.pdf) Those values pull the device out of the realm of vague “natural interface” language: somebody had to size the links, balance the mechanism, choose sensors and decide what body the controller should fit.

Around it sits a sizeable simulator. The research programme built 19 virtual environments, 15 for training and four for evaluation, with tasks including digging, dumping and moving debris, logs and rocks.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488) Unity handles much of the scene and physical interaction; a real-time control layer talks to the WSI and the traditional joystick setup.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488)[3](https://construction-robots.github.io/papers/72.pdf)

The team was therefore testing more than whether somebody could make an arm move in the expected direction. They tried to reproduce enough of an excavator course to watch people work through repeated tasks and increasing difficulty.

## Thirty-seven percent

The strongest result is not the catchy version about novices becoming experts overnight.

At first use, **average cycle time with the WSI was 37 percent shorter than with the joystick interface**.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488) The WSI's pre-training performance was also close to the level novices reached after training on joysticks.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488)

The earlier 2025 experiment puts numbers around that comparison: twelve people with no excavator experience were split into two groups of six, one using the WSI and the other joysticks, then trained for seven days.[3](https://construction-robots.github.io/papers/72.pdf) Over that week the joystick group cut average cycle time by **46 percent**, while the WSI group changed by only about one percent because it had already started near the later performance level.[3](https://construction-robots.github.io/papers/72.pdf)

> **Read the result carefully**
> - shorter first-use cycle time with WSI than joysticks: 37%
> - training period in the precursor protocol: 7 days
> - joystick cycle-time improvement after training: 46%
> Sources: Journal of Computing in Civil Engineering 2026; 2025 workshop paper

This is why the interface is more interesting than another training aid: instead of coaching people through the old control mapping more efficiently, it changes the mapping so less translation has to be learned in the first place.

There is still plenty to learn about excavation. The point is that memorising a control convention need not consume quite so much of the beginner's attention.

## Not experts

MIT News describes novices on the WSI as “just as good as experts from the start”.[1](https://news.mit.edu/2026/mit-engineers-design-better-controller-operating-construction-diggers-0820) Read outside the tested tasks, that sentence becomes much bigger than the evidence underneath it.

The final paper itself is more careful. On some simple tasks, such as digging and dumping, trained novices using joysticks eventually showed no significant performance difference from experts.[2](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488) That does not turn a laboratory participant into somebody ready for a live site full of poor visibility, changing ground, other workers and expensive consequences.

The precursor also exposes the narrow human sample behind the early result: **12 novices**, all healthy men aged 24 to 35, plus nine expert operators who were asked about simulator realism. The authors noted that they had not yet made a quantitative comparison between those experts’ simulator performance and their performance in a real excavator.[3](https://construction-robots.github.io/papers/72.pdf)

There is another small warning inside the hardware description: dimensions were chosen using male anthropometric data, and the springs compensated for an estimate of an average man's arm weight.[3](https://construction-robots.github.io/papers/72.pdf) The controller may feel intuitive, but “intuitive” still has dimensions, reach and a body hidden inside the word.

That does not invalidate the interface. It simply means a universal claim has to wait for a broader set of users and real machines.

## Missing weight

The other large gap is something operators notice immediately in physical machinery: the machine pushes back.

A real excavator communicates through vibration, sound, hydraulic load, soil resistance and movement of the chassis; the WSI studied here mostly gives the operator a geometric relationship plus the visual response of the simulator, leaving much of the force information from a bucket biting into material outside the loop.

The team is working on haptic feedback so the small arm can resist the operator when the virtual machine encounters a load.[1](https://news.mit.edu/2026/mit-engineers-design-better-controller-operating-construction-diggers-0820)[3](https://construction-robots.github.io/papers/72.pdf) Once force enters the loop, the controller has to communicate the difference between empty air, loose sand and a stubborn obstacle without becoming tiring or unstable; that problem sits much closer to the craft of real digging than a visual animation does.

> **What is still missing**
> - Direction, linkage geometry, bucket position and coordination.: Already mapped
> - Force, weight, soil resistance, vibration and other cues from a real machine.: Still partial
> - Reduce mental translation before training begins.: Promise
> - Add haptics without making the interface harder to use.: Next step
> Visual and kinematic intuition are only part of an operator's craft.

Copying the linkage geometry was the first simplification. Reproducing the relationship between a heavy bucket and ground that changes from one scoop to the next is harder, and probably more revealing about where expert judgement really lives.

## Useful complexity

Construction robotics is often described as a path towards taking the operator out of the loop: assisted motion, remote control, semi-autonomous tasks, eventually more autonomy. The WSI asks a different question before any of that happens.

Why should the human still perform a nonlinear coordinate conversion that software can do perfectly well?

A patent application published this summer describes the wider idea of machine controls whose geometry represents the mechanism being controlled.[4](https://patents.justia.com/patent/20260225249) Excavators are an unusually legible example, but the same design problem appears in industrial arms, forestry equipment, cranes and other machinery where an operator's spatial intention is translated into several abstract axes.

This separates two kinds of expertise that are easy to confuse: a good excavator operator reads the ground, judges load, chooses where to position the machine, notices risk and avoids wasting motion, all of which belong to excavation itself. Remembering which pair of joystick movements produces a diagonal bucket path belongs much more to the interface convention chosen decades ago.

The miniature controller does not make the machine simple, and it certainly does not make the job simple. What it does is identify one piece of complexity that belonged to the controls and hand that piece back to the computer.

## References

1. [Jennifer Chu, MIT engineers design a better controller for operating construction diggers, MIT News, August 20 2026](https://news.mit.edu/2026/mit-engineers-design-better-controller-operating-construction-diggers-0820)
2. [H. Ishida et al., A Comprehensive Training Platform for Excavator Operators, Journal of Computing in Civil Engineering, 2026](https://ascelibrary.org/doi/10.1061/JCCEE5.CPENG-7488)
3. [H. Ishida et al., Joint-Space or World-Space Interfaces: Examining the Performance of Novice Excavator Operators, 2025](https://construction-robots.github.io/papers/72.pdf)
4. [U.S. Patent Application 20260225249, Control Methods and Interfaces for Controlling Machines](https://patents.justia.com/patent/20260225249)