Open source · GitHub Digital Future Lab · UHasselt Flanders Make Space teleoperation · reproducibility
UNITE · modular Unity testbed
Reproducible delayed space teleoperation
A shared, configurable apparatus for studies of driving robots under communication delay. Grounded in a scoping review of 83 studies.

Robots on the Moon or Mars are driven with seconds to minutes of round-trip delay. Researchers test interfaces that help operators cope, but each study builds its own setup, and papers rarely describe it completely enough to rebuild. UNITE turns the whole setup (delay, input, feedback, vehicle, assistance, task, logging) into a configuration you can share, rerun, and change one factor at a time.

Dries Cardinaels · Thomas Pietrzak · Raf Ramakers · Kris Luyten Digital Future Lab · UHasselt · Flanders Make · Diepenbeek, BE  |  CRIStAL · Univ. Lille, Inria, CNRS · Lille, FR
83 studies reviewed 7 design requirements 13 module contracts 1 study reconstructed
The gap

Every delay study builds its own rig

Delay slows operators, raises error rates, and pushes them into a move-and-wait strategy. Techniques such as predictive displays and shared control help, but they are evaluated on study-specific apparatus. Comparing two techniques, or rerunning one, means reconstructing that apparatus from prose.

What papers report

A summary of the setup

  • "2.56 s delay", without saying how it splits between command and feedback paths
  • "Keyboard control", without the key-to-speed mapping or hold/toggle behaviour
  • "Realistic physics", without steering geometry or model parameters
UNITE
What UNITE stores

The setup itself

  • Delay per stream and direction, with its temporal behaviour and packet loss
  • Explicit device-to-command mappings, including scaling
  • A vehicle model with steering geometry as a named field, saved as a shareable Unity project

Existing tools cover parts: Gazebo and CARLA simulate robots and vehicles, TELECARLA adds networked teleoperation, PsychoPy and UXF run trials. None represents the full delayed-teleoperation loop as one configuration.

Contributions

Review, testbed, reconstruction

C1

Scoping review of 83 studies

Charts how delayed human-in-the-loop mobile teleoperation experiments are built and reported, and where the reports leave gaps.

C2

UNITE, an open testbed

A modular Unity environment where communication, input, feedback, vehicle, assistance, task and logging are separately configurable components.

C3

A reconstruction

A prior lunar-delay user study rebuilt in UNITE without extending its architecture, exposing three apparatus details the original paper left open.

Try it · interactive

Drive a rover through the loop

A miniature of UNITE's teleoperation loop. Your input is sampled at 10 Hz, queued on the uplink, applied by a vehicle model, observed, and queued on the downlink before you see it. The solid rover is what the operator sees. Change the delay split, switch assistance modules on and off, and watch the configuration on the right update.

Condition preset
Click here, then drive with arrow keys or WASD
Communication · DR1
Uplink (commands)2.56 s
Downlink (vehicle state)0.00 s
Round trip2.56 s
Assistance · DR5
Predicted path (operator side)
Route keeping (remote side)
Vehicle · DR4  /  View
Terrain disturbance
Show true rover (hidden from operator)
Touch controls
Trial log · DR6 / DR7
Waiting for first command.
This condition as a shareable configuration (live)

    

Illustrative simulation written for this poster, not UNITE's code. The course, vehicle model, disturbance and route-keeping rule are simplified stand-ins; the configuration format is a sketch of the idea, not UNITE's file format. Try: turn on the predicted path, then enable terrain disturbance and see prediction and truth drift apart.

Same round trip, different loop

Where does the delay sit?

A study that reports only "2.56 s round trip" leaves the split open. For the operator the response always arrives 2.56 s after a key press, but the split decides how stale the displayed image is and how fresh the state is that remote-side assistance acts on. Maag et al. ran 200/0, 0/200 and 100/100 ms splits as separate conditions.

all uplinkdownlink shareall downlink
Split assumed in the reconstruction

Fixed 2.56 s round trip, as in the reconstructed study. Diagram only; it ignores processing and frame time.

Scoping review

From 918 records to 83 studies

ACM DL, IEEE Xplore, Scopus and Web of Science, up to and including 2025. A study was included when a human operated a ground or surface vehicle, delay was an experimental condition, and there was a real user study (procedure, participants, participant-level results, or operator count). Fifteen method demonstrations with a human in the loop but no user study were removed during charting.

IDENTIFIED

0 records

From four databases; 6 more found by manual search.

SCREENED

0 unique

After automatic (ASySD) and manual deduplication.

INCLUDED

0 studies

Charted with a codebook built from a 10-study pilot, following stages of the teleoperation loop.

DERIVED

0 requirements

Each backed by variation the charting found across studies.

Unreported details were coded as "not reported" and never inferred from hardware descriptions. Participant counts, training and subjective measures were charted but left to experiment-management tools: UNITE covers the apparatus, not the study protocol.

Design requirements

Seven things a reusable setup must capture Tap to expand.

Architecture

The loop as swappable modules

A staged orchestrator runs the loop with separate uplink and downlink paths. Thirteen modules, grouped below, each declare what they consume and publish. Implementations are interchangeable when their contracts match.

CONTRACTS

Boundaries are explicit

Each module states what it does not do. Communication treats payloads as opaque; command mapping expresses intent but never moves the vehicle.

ASSISTANCE

Three attachment points

Before uplink transmission, after uplink reception on the robot, and after downlink reception at the operator. The same shared-control function can be placed on either side as separate conditions.

OBSERVATION

Two refreshes per step

Remote state is captured before remote-side assistance (available locally, undelayed) and again after the vehicle update (optionally sent down the delayed link).

Demonstration

Rebuilding a CHI '26 lunar-delay study

Cardinaels et al. (CHI '26) had 24 participants drive a simulated lunar rover under a fixed 2.56 s round trip, comparing three feedforward visualizations (Network, Path, Envelope) with a delayed-video baseline. It was not in the review corpus. All four conditions became scenes that differ only in the downlink operator-side assistance module; no architectural extension was needed. Three details still had to be decided:

DR1 · DELAY SPLIT

Uplink or downlink?

Only the round trip was reported. The reconstruction assigns all 2.56 s to the uplink and 0 s to the downlink.

DR4 · DISTURBANCE

Parameter values

Six disturbance sources were described, with Perlin noise for some, but values were "iteratively calibrated". Values in UNITE are explicit choices meant to give comparable variation, not the original calibration.

DR5 · ROUTE KEEPING

Terrain or assistance?

The paper credits cliffs and boundaries. The original source shows a slope-based boundary detector with a threshold and reverse-on-contact, now ported as a remote-side assistance module. A reader of the paper alone could not recover this.

The original authors overlap with this paper, so source code was available as a reference. That is what made gaps 2 and 3 resolvable at all.

Honest limits

What this does not show yet

L1

One reviewer charted the corpus

Mitigated with a fixed codebook, a pilot pass, and a separate "unclear" code that was resolved against field definitions.

L2

One reconstruction

Initial evidence only, done by an overlapping author team. Other vehicles, techniques, communication conditions and tasks remain to be tested.

L3

Geometry is not a named factor

Values stored on components (e.g. lighting) are recoverable from a scene; properties that emerge from geometry (e.g. corridor width) are not labelled as factors and still need reporting.

L4

Apparatus, not protocol

Participants, training, counterbalancing and questionnaires are out of scope and left to tools such as UXF or PsychoPy.

In one breath

Share the setup as a configuration, not a paragraph

A review of 83 delayed-teleoperation studies yields seven parts every setup has. UNITE makes each part a configurable module, so a study can be rerun, compared under matched conditions, or changed in one factor. Rebuilding a carefully reported study still surfaced three unstated apparatus choices, which UNITE turns into explicit configuration fields.