In a field tent at night, two technicians inspect a mud-caked quadruped robot with a leg panel removed.

The proving ground · For robotics teams and researchers

Bring us something you think can't be broken.

Robotics teams spend enormous effort teaching machines to succeed. MechaSafari is being designed to discover, systematically, how they fail around real, motivated, unpredictable people.

Why humans

Labs test what you thought of. People test what you didn't.

Structured testing is essential, and it is bounded by the imagination of whoever wrote the test plan. Give a motivated person a goal and a machine to beat and they will do things no plan contains: split up, double back, go completely still, go loud, use the weather, use each other.

That is the behaviour embodied systems will meet in the world. MechaSafari is an environment built to produce it on purpose, safely and repeatedly, with every instrument running.

What we're designing to test

  1. 01

    Adversarial field testing

    Your platform against human teams with a goal and a clock.

  2. 02

    Navigation robustness

    Mud, snow, deadfall, slope, water, darkness and fog.

  3. 03

    Perception edge cases

    Occlusion, low sun, thermal crossover, rain on the lens, people who don't behave like the training set.

  4. 04

    Human-machine interaction

    How people read, approach, avoid and mislead your system, and how it responds.

  5. 05

    Fail-safe validation

    Does it stop when it should, every time, under pressure and in bad conditions?

  6. 06

    Containment testing

    Geofence, boundary and recall behaviour at the edges of the field.

  7. 07

    Cross-platform benchmarking

    The same scenario and conditions, run against different machines.

  8. 08

    Environmental robustness

    Cold soak, wet, night and the change of seasons.

  9. 09

    Recovery behaviour

    Falls, stalls, lost localisation and lost comms, and what happens next.

  10. 10

    Independent reproducibility

    A finding counts when it happens again on demand, observed by someone who didn't build the machine.

A robot's forward camera view up a mossy slope, a participant climbing away between the trees.

How an engagement would work

  1. Scope

    You bring a platform and the questions you care about. We agree what is in bounds, what isn't, and who owns what.

  2. Configure

    We build scenarios around those questions: terrain, light, weather, team size, objective.

  3. Run

    Instrumented runs with participant teams: synchronised multi-angle video, thermal coverage, participant positions, a ground-truth map, and the machine logs you choose to share.

  4. Validate

    Anything interesting is re-run under controlled conditions until it reproduces, or doesn't.

  5. Report

    Findings, reproductions and the evidence behind them, under terms agreed before the first run.

Data, credit and IP

Your platform's data is governed by the agreement you sign, not by our convenience.

Participants' behaviour will not be sold as anyone's training corpus. When a person originates a genuine discovery about your system, it is recorded as theirs: credited and, where it creates value, shared. We think that makes the findings more trustworthy, not less.

What we won't do

Where we are, honestly

Exists today

The concept, the scenario design, and this site.

Not yet

A field, a fleet, a dataset, customers, partners or certifications.

Why talk now

Early partners shape the test protocols and the first scenarios.

Talk to the proving ground.

Start the conversation