A robot used step-by-step reasoning to complete 91.67% of 240 trials

The system generated action plans by separating visual input, language, reasoning and movement instead of relying on task-specific demonstrations or policy training.

Editorial illustration — not from the study.

Researchers developed a robot system that separates seeing, understanding language, reasoning and acting, then connects those functions with numerical signals. A large language model supplies common-sense, mathematical and physical reasoning to produce an action plan through explicit logic.

The system did not use task-specific demonstration data or manipulation-policy training. In six experimental suites covering scene understanding, manipulation, language interaction, complex reasoning and long sequences of actions, it succeeded in 91.67% of 240 real-world trials.

How the robot performed

The researchers presented a robot system that decouples vision, language, reasoning and action. These modules are linked through numerical signals, while a large language model uses common-sense, mathematical and physical reasoning to generate a complete action plan through explicit logic.

The system required neither task-specific demonstration data nor manipulation-policy training. Researchers evaluated it in six experimental suites covering 3D scene understanding, qualitative and quantitative manipulation, language interaction, complex reasoning and long-horizon multi-step autonomy. Across 240 real-world manipulation trials, the robot achieved a 91.67% success rate. The researchers also report that the robot’s behaviors were interpretable within this approach.

Why the approach matters

The work suggests that explicit task reasoning can support a range of robot-manipulation tasks without collecting task-specific demonstrations or training a manipulation policy for each one. By separating the system’s functions and generating actions through stated reasoning steps, the approach also aims to make robot behavior easier to interpret. The results provide evidence for this approach in the study’s six experimental suites, while leaving its performance in other settings to be tested.

Evidence and open questions

The evidence comes from six experimental suites and 240 real-world manipulation trials. The abstract does not describe the individual tasks, provide comparison results against other systems, or establish how the robot would perform outside these designed tests. The study also reports results for this particular system and experimental setup, so broader reliability and generalization remain unclear.

Reliable AI partnerships depend on people, not just stronger models

A survey of human-AI collaboration argues that more capable AI systems do not automatically produce reliable partnerships. It identifies human-centered design, governance and attention to high-stakes uses as central to making these systems trustworthy and beneficial.

9 hours ago·6 views

AI & Computing

RNA splicing tools can follow irrelevant clues and miss RNA’s shape

Deep-learning tools for predicting RNA splicing can rely on genomic clues unrelated to splicing and miss important effects of RNA structure, a study finds. These weaknesses were linked to systematic errors, especially when the tools evaluated sequences that differed from those used for training.

20 hours ago·5 views

AI & Computing

Better-trained AI finds more flaws in unfamiliar software projects

AI vulnerability detectors often perform less well on software projects outside their training data. A study finds that cleaner, more varied training data and an encoder-based design improved detection, including a reported 6.8% gain in recall on the BigVul benchmark.

2 days ago·34 views