How good are the most powerful artificial intelligence systems at controlling robots and can they cope with ordinary everyday tasks such as clearing a table or arranging objects? A new platform, jointly developed by technology giant NVIDIA and the Bulgarian INSAIT Institute at Sofia University “St. Kliment Ohridski“, provides an answer to these questions. It is called LADDER and allows for the evaluation and comparison of the capabilities of leading AI models to control robots when performing a series of sequential actions.
Although artificial intelligence already solves complex mathematical problems, and humanoid robots can perform impressive movements, controlling robots during ordinary everyday activities remains a serious scientific challenge. Moving an object is a relatively easy task, but when a robot has to move several things in succession, the probability of errors increases significantly.
The development is aimed at one of the most promising areas in modern technology - physical artificial intelligence (Physical AI), in which AI systems must not only process information, but also interact with the real environment.
NVIDIA is one of the world's technology giants, whose developments are at the heart of the revolution in artificial intelligence. The company is among the leading factors in the development of the most powerful AI models, and robotics and physical artificial intelligence are among its strategic priorities. It is in this area that the development of the Bulgarian INSAIT attracts the attention of NVIDIA, which recognizes its potential and is involved in its development. The result of the joint work is the LADDER platform, which has the potential to become an important reference point for assessing progress in robotics and the development of the next generations of intelligent robots. The collaboration is a significant recognition of INSAIT's scientific potential and its place among international research centers.
The initial results show that even the most powerful systems are still far from reliably controlling robots. Among the leading models in the ranking are GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon. Although they perform better than the others, they fail to complete a significant part of the tasks. In the most complex tests, involving eight to ten consecutive actions, even the best models achieve success only in about a third of the cases. Even more serious difficulties are experienced by models developed specifically for controlling robots. One of the leading ones manages to move a single object, but almost never copes with five or more consecutive actions.
The study also identifies another significant challenge - the high cost of using the most powerful AI models. In some of them, the average value of a successfully completed task exceeds $6, and a full evaluation of the model through nearly 2,000 trials would cost about $8,000.
The platform is publicly available and will be updated with the results of new AI models. The ranking, videos of the task completion, and the full scientific study can be viewed on the website https://ladderrobotics.ai.