IROS 2026 · Pittsburgh · Sept 27 — Oct 1

Real-world
AI vs Human
navigation, in the wild.

A globally distributed competition pitting autonomous policies against expert human teleoperators. Same rovers, same missions, across 10+ cities.

Register your team → See the brief
Platform
Earth Rover Mini+
Tracks
4 · cross-domain
Edition
Round 03
Cost
Free to enter
IROS 2024 · Abu Dhabi
Same fleet Same missions 10+ cities Open datasets · 2K + 7K hours Open-source SDK
The challenge

A globally distributed AI vs human navigation competition. Same fleet, same missions, 14 sites worldwide.

Global fleet · 14 sites
Sites14
Outdoor rovers35
Indoor / off-road5
Marathon (New)03
01

The gap is real-world.

State-of-the-art navigation lives in sim and curated datasets. We test it where it actually matters.

02

One fleet, many cities.

A shared, calibrated rover fleet across 14 sites worldwide. Same hardware, same missions, same network.

03

AI vs human ceiling.

IROS 2024: top AI 15.4 vs human 42.0. ICRA 2025: 23.78 vs 30.0. The gap is closing fast.

04

Pittsburgh, on the day.

Final round at IROS 2026 — local missions on the venue floor, results scored against the global field.

Tracks

Four tracks. One model to rule them all.

Mission scoring = difficulty × time. Highest aggregate wins.

01
Urban · GPS

Urban navigation

Sidewalks, crossings, pedestrians. GPS waypoints across multiple cities.

02
Indoor · Image

Indoor navigation

Hallways, atriums, labs. Image-goal driven — no GPS.

03
Off-road · Image

Off-road terrain

Trails, slopes, gravel. Unstructured natural terrain.

04
New · Endurance

The marathon

Cross-domain endurance — urban → indoor → off-road, single policy.

01 · Mission

Reach the checkpoint

Within 15 m GPS tolerance. Each mission scores its difficulty rating (1–10).

02 · Score

Aggregate points

Sum of difficulty ratings across all completed missions. Highest aggregate wins.

03 · Tiebreaker

Time on shared missions

Lowest cumulative completion time across shared missions wins ties.

Resources

Dataset, SDK and real-world evaluation — free for all participating teams, released ahead of the venue.

01 · Datasets

FrodoBots‑2K + Berkeley‑FrodoBots‑7K

Two open datasets of in-the-wild human-teleoperated drives across 10+ cities. FrodoBots-2K (~2,000 hours) is the curated subset used in the MBRA paper. Berkeley-FrodoBots-7K (~7,000 hours) is the full set, reannotated with MBRA for higher-fidelity action labels.

02 · SDK

Remote Access SDK

Host your model on your own compute. Stream front camera in, control commands out. GPS + map exposed.

View on GitHub →
03 · Real-world eval

Real-world eval at remote sites

Up to 20 hours/week of pre-event testing across multiple cities. Bot-walker support, identical to venue conditions.

Allocations open with registration →
Hardware

Earth Rover Mini+ — a standardized 4WD teleop chassis with cameras, GPS, IMU and 4G. All compute is off-board.

Earth Rover Mini+

Standardized competition rover

Spec sheet

Drive
4WD · turn-in-place
Cameras
Front + Rear
Localization
GPS · IMU
Comms
4G LTE
Compute
Off-board · no limit
Stream rate
~20 Hz · ~500 ms latency
Action space
Fwd · Bwd · L · R
Fleet
10+ cities
Schedule

Two days of competition in Pittsburgh, Sept 30 — Oct 1, 2026, at the David L. Lawrence Convention Center.

Day 01 · Sept 30

Practice + Rounds 1–4

Remote urban · off-road · indoor
  • 09:00Connectivity & model testing
  • 10:00Round 01 — Remote Urban
  • 11:00Round 02 — Remote Urban
  • 12:00Lunch
  • 14:00Round 03 — Off-Road
  • 15:00Round 04 — Indoor
Day 02 · Oct 1

Final rounds + awards

Local missions · trophy
  • 09:00Connectivity recheck
  • 10:00Round 05 — Remote Urban
  • 11:00Round 06 — Remote Urban
  • 12:00Lunch
  • 14:00Final round — Local missions, IROS venue
  • 16:00Awards & wrap-up
Benchmark

AI is closing the gap against humans.

Top AI score per edition, as a fraction of the all-time top human run. The bar fills toward the human ceiling.

Previous edition · highlights
All-time human ceiling 100%

IROS 2026 · Pittsburgh

Open · Oct 2026
TBD
Your model goes here.

ICRA 2025 · Atlanta

May 2025
57%
Top AI: National University of Singapore — closing in on the human ceiling.

IROS 2024 · Abu Dhabi

Oct 2024
37%
Top AI: Seoul National University. Top human (masterchi_) set the ceiling this edition.
Co-organizers

FrodoBots, NUS, GMU, NYU, Princeton, Google DeepMind, Meta AI, Georgia Tech and UMD.

Chen Feng
Chen Feng
New York University
Xuesu Xiao
Xuesu Xiao
George Mason University
Michael Cho
Michael Cho
FrodoBots
David Hsu
David Hsu
National University of Singapore
Dhruv Shah
Dhruv Shah
Princeton University
Joanne Truong
Joanne Truong
Meta AI
Ted Xiao
Ted Xiao
Google DeepMind
Naoki Yokoyama
Naoki Yokoyama
Georgia Tech
Wenhao Yu
Wenhao Yu
Google DeepMind
Tingnan Zhang
Tingnan Zhang
Google DeepMind
Dinesh Manocha
Dinesh Manocha
University of Maryland
Santiago Pravisani
Santiago Pravisani
FrodoBots
Made possible by
Main Organizer
Berkeley → Stanford · ~50 mi
Track 04 · New for 2026

The marathon.

A long-form endurance test for navigation policies. One long-distance run strung together from mixed legs across urban arterials, off-road terrain, and indoor handoffs — single policy end-to-end.

01 · Format

One long run · seven legs

The run chains together urban, off-road and indoor segments. No driver hand-offs, no policy swaps mid-run. The route probes whether one model can hold up across domain shifts and accumulate kilometers without drift.

02 · Route

UC Berkeley → Stanford

~50 mi over the Bay Area corridor, split into seven named legs (1.1, 1.2, 2.1, 2.2, 3.1, 3.1.5, 3.2). Mix of city sidewalks, bay-side bike paths, overpasses, and suburban arterials. The flyover at left traces the planned path.

03 · Why it matters

Generalization, under load

Single-domain benchmarks don't reveal whether a policy can handle the seams. The marathon stresses long-horizon recovery, scene-shift robustness, and the cumulative cost of small mistakes — the failure modes researchers most often miss in sim.

04 · Scoring

Per-leg difficulty · TBD

Leg-level scoring formula being finalized with FrodoBots, GMU and NYU. Expected to follow the same 1–10 difficulty rubric as the other three tracks, with cumulative time as the tiebreaker.
Earth Rover navigating real-world city sidewalk
Real-world conditions
GPS checkpoints across a city route
GPS waypoint checkpoints
Track 01 · Flagship

Urban navigation.

Outdoor city sidewalks across multiple cities worldwide. Goals are GPS waypoints; the rover must reach within 15 m of each checkpoint, in real traffic, with real pedestrians, on real sidewalks.

01 · Goal

GPS waypoints · 15 m tolerance

Each mission is a sequence of GPS checkpoints. A checkpoint counts as reached when the rover comes within ~15 m — the tolerance accounts for GPS noise on consumer hardware in dense urban canyons.

02 · Real-world variability

Same fleet, different cities

Sites span 10+ cities globally — Botswana, Mauritius, Kenya, Nigeria, Mexico, Costa Rica, Panama, the US. Sidewalk widths, curb cuts, crossing geometry, traffic mix, lighting and weather all vary. The policy that wins is the one that generalizes.

03 · Difficulty rubric

Rated 1–10 per mission

Difficulty is set per mission based on road crossings, pedestrian and traffic complexity, expected drive time, number of turns, and path type — parks and rural roads are easier; bike paths, sidewalks and on-road segments scale up.