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.

63.3%

The best AI model still fails often

GPT-6 Astra, the best model we tested, finishes 63.3% of the tasks. It fails on about 1 in 3.

$6.49

The best is also the most expensive

GPT-6 Astra costs $6.49 per finished task, 20 times more than the runner-up.

42.8%

Cheaper models fail more

Gemini 3.7 Flash costs $0.33 per finished task, but finishes only 42.8% of them.

Why this matters

2× speed

“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.

63.3%of tasks finished by the best AI model we tested
34.8%when a task has 8 to 10 objects to move

How LADDER works

Short task“Put the banana in the bin”
Long task“Put the banana, both oranges, the lemon, the lime, the avocado, the onion, the lychee, the pumpkin and the pomegranate in the bin”
4× speed

Same robot, same table, same kind of instruction. Each extra object is one more chance to fail, so long tasks show how far an AI model really gets.

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

0%25%50%75%100%12–34–56–78–10objects to move in one taskGPT-6 Astra (GPT-Policy): 92.0%GPT-6 Astra (GPT-Policy): 83.2%GPT-6 Astra (GPT-Policy): 62.6%GPT-6 Astra (GPT-Policy): 44.1%GPT-6 Astra (GPT-Policy): 34.8%Gemini 3.7 Flash (PSP-Agent): 80.9%Gemini 3.7 Flash (PSP-Agent): 63.6%Gemini 3.7 Flash (PSP-Agent): 40.3%Gemini 3.7 Flash (PSP-Agent): 24.0%Gemini 3.7 Flash (PSP-Agent): 8.9%Robot model alone: 87.2%Robot model alone: 21.1%Robot model alone: 3.5%Robot model alone: 2.3%Robot model alone: 0.8%0%25%50%75%100%12–34–56–78–10objects to move in one taskGPT-6 Astra (GPT-Policy): 92.0%GPT-6 Astra (GPT-Policy): 83.2%GPT-6 Astra (GPT-Policy): 62.6%GPT-6 Astra (GPT-Policy): 44.1%GPT-6 Astra (GPT-Policy): 34.8%Gemini 3.7 Flash (PSP-Agent): 80.9%Gemini 3.7 Flash (PSP-Agent): 63.6%Gemini 3.7 Flash (PSP-Agent): 40.3%Gemini 3.7 Flash (PSP-Agent): 24.0%Gemini 3.7 Flash (PSP-Agent): 8.9%Robot model alone: 87.2%Robot model alone: 21.1%Robot model alone: 3.5%Robot model alone: 2.3%Robot model alone: 0.8%
GPT-6 Astra (GPT-Policy)Gemini 3.7 Flash (PSP-Agent)Robot model alone

Share of tasks finished. Every model is on the leaderboard, under Details.

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.

2× speed

“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.

0%25%50%75%100%Robot model alone, simulation: 19.6%Proprioceptive verifier, simulation: 25.6%Gemini 3.7 Flash via PSP-Agent, simulation: 51.4%SimulationRobot model alone, real robot: 5.6%Proprioceptive verifier, real robot: 58.3%Gemini 3.7 Flash via PSP-Agent, real robot: 75.0%RealRobot model alone, simulation: 9.8%Proprioceptive verifier, simulation: 24.7%Gemini 3.7 Flash via PSP-Agent, simulation: 38.5%SimulationRobot model alone, real robot: 15.7%Proprioceptive verifier, real robot: 40.7%Gemini 3.7 Flash via PSP-Agent, real robot: 81.5%RealFrankaWidowX0%25%50%75%100%Robot model alone, simulation: 19.6%Proprioceptive verifier, simulation: 25.6%Gemini 3.7 Flash via PSP-Agent, simulation: 51.4%SimulationRobot model alone, real robot: 5.6%Proprioceptive verifier, real robot: 58.3%Gemini 3.7 Flash via PSP-Agent, real robot: 75.0%RealRobot model alone, simulation: 9.8%Proprioceptive verifier, simulation: 24.7%Gemini 3.7 Flash via PSP-Agent, simulation: 38.5%SimulationRobot model alone, real robot: 15.7%Proprioceptive verifier, real robot: 40.7%Gemini 3.7 Flash via PSP-Agent, real robot: 81.5%RealFrankaWidowX
Robot model aloneProprioceptive verifierGemini 3.7 Flash via PSP-Agent

Get involved

Add your model to the leaderboard Run the tasks with our free code and send us the results. We check every score.
Want to learn more? Technical details: full results, the list of tasks, tests on real robots, every video, the paper and the code.
Paper Code Hugging Face

Team

Petko Petkov1, Giuliano Albanese1, Nikolay Nikolov1, Luc Van Gool1, Jonathan Tremblay2,†, Danda Pani Paudel1,†
1Institute for Computer Science, Artificial Intelligence and Technology 2NVIDIA †Equal supervision
INSAIT NVIDIA

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}
}