How well can today’s best AI models control a robot?
Success vs cost
Each dot is one AI model. Higher means it finishes more tasks. Further right means it costs more. Models marked Lite are scored on the 10 LADDER-Lite tasks.
Success comes at a price
GPT-6 Astra costs $6.49 per finished task, 20 times more than Gemini 3.7 Flash via PSP-Agent, which finishes 42.8% of the tasks.
New models on LADDER-Lite
On LADDER-Lite, Gemini 4 Argon finishes 55.9% at $1.33 per finished task; Claude Opus 5.5 finishes 49.0% at $7.02 per finished task. GPT-6 Astra finishes 62.5% at $6.62.
Why this matters
“Put the banana, both oranges, the lemon, the lime, the avocado, the onion, the lychee, the pumpkin and the pomegranate in the bin”
Looks easy? Not for robots.
Putting ten fruits and vegetables in a bin is child’s play for a person. For a robot it is still very hard: it has to find each object, pick it up, carry it and put it down, ten times in a row, without a single mistake.
Robots that can reliably handle everyday objects are one of the biggest unsolved challenges in AI. LADDER measures how far today’s best AI models have come.
How LADDER works
Short task · 3 objects
“Put the lemon and the lime on the plate and put the orange in the bin”
Long task · 9 objects
“Put the banana, the lemon and the lime on the plate and put both oranges, the avocado, the onion, the lychee and the pumpkin in the bin”
The same table and the same two containers. The long task only adds more objects to sort, so any drop in success comes from the length alone. One slip anywhere and the whole task fails. Both clips are GPT-Policy, the best system we tested.
145 everyday tasks
From 1 to 10 objects: putting things on a plate or in a bin, in a set order, moving them from one place to another, or leaving some where they are.
Checked automatically
Every task runs in a robot simulator that checks where each object ends up. A task only counts when everything is in the right place.
The longer the task, the more often robots fail
Share of tasks finished, the three simulators weighed equally. Hover the chart to read the values. Click a name to show or hide it.
Simulation matches the real world
We ran the same kinds of tasks on two real robots. The models rank in the same order as in simulation.
“First put the blue cup in the bowl, then put the carrot in the bowl, then put the eggplant in the bowl”
A real Franka robot in our lab, controlled by Gemini 3.7 Flash via PSP-Agent.
Team
Citation
@article{ladder2026,
title = {{LADDER}: A Long-Horizon Robot Manipulation Benchmark},
author = {Petkov, Petko and Albanese, Giuliano and Nikolov, Nikolay and Van Gool, Luc and Tremblay, Jonathan and Paudel, Danda Pani},
journal = {arXiv preprint arXiv:2610.18053},
year = {2026},
url = {https://arxiv.org/abs/2610.18053}
}