Claude Opus 5.5 robot programs achieve 92.1 percent success rate
A new study shows that AI-generated robot programs can complete physical tasks with high reliability without any human intervention during execution.
The short version
- The AI model achieved a 92.1 percent success rate across eight different physical tasks.
- Programs reached 100 percent success when developed using computer simulations.
- No human or active AI agents intervened during the actual execution of the robot tasks.
- The hardest task was packing bread, which had a lower success rate due to the object's deformable nature.
Why it matters
This progress demonstrates that AI can move from digital reasoning to reliable physical action. It shows that using simulations to train these models can make robot deployment faster and more accurate.
Reliability in physical tasks
A preliminary study released on November 12, 2026, examined the performance of robot programs generated by the Claude Opus 5.5 AI model. The research tested the systems on eight distinct physical tasks to see how reliably they could perform without constant supervision. The results showed that the programs mastered 221 out of 240 total runs.
One of the most significant aspects of this experiment was the lack of intervention during the execution phase. Neither a human nor an active AI agent was present in the control loop while the robots were performing their assigned duties. All support was provided exclusively during the initial development phase before the programs were set in stone.
The study utilized two Agilex-PIPERX robot arms and three Orbbec-DaBai-DC1 cameras to capture the visual environment. These tools allowed the researchers to measure how well the AI-generated instructions translated into physical movements. The success rate of 92.1 percent suggests a high level of autonomy for these types of robotic programs.
Performance across different tasks
The researchers tested a variety of tasks ranging in difficulty to see where the AI excelled and where it struggled. For example, the system achieved a perfect success rate of 30 out of 30 runs when inserting charging plugs. Other tasks, such as stacking cups and rearranging packages, were also highly successful with 28 successes each.
More complex movements showed slightly lower success rates but still remained highly consistent. Opening bottles, inserting flowers, and packing shoes each resulted in 27 successful runs. These tasks required the robot to interact with objects in more varied ways than the simpler assembly tasks.
The most challenging task for the system was packing bread, which resulted in only 25 successes. The authors attributed this lower performance to the nature of the object itself. Because bread is a deformable object, it presented a greater challenge for the robot to handle consistently compared to rigid items.
The role of simulation
The study also compared how the use of computer simulations during the development phase affected the final results. When a simulation environment was used to help create the programs, the success rate jumped to 100 percent. This method also proved to be more efficient in terms of both time and computational resources.
In contrast, developing programs without a simulation environment led to a lower success rate of 86.7 percent. This lack of simulation also increased the development time to 3.7 hours. Using a simulation allowed the team to reach the 100 percent success mark in just 2.6 hours while consuming 1.09 million output tokens.
During the development process, the researchers allowed for up to 10 human-assisted scene resets per task. However, these resets were strictly for the training and setup phase. The data suggests that simulation is a powerful tool for refining AI instructions before they are deployed on physical hardware.
Research and future questions
The research involved several prominent figures, including Ken Goldberg from the University of California, Berkeley, as well as Fei-Fei Li and Jiajun Wu from the Stanford University. Their work provides a look into the current state of autonomous systems and the progress of AI-generated robotics. The study highlights the transition toward more reliable, independent robot operations.
Despite the positive results, the authors noted several areas where more research is still needed. One major gap is the lack of a direct comparison between these programs and control methods that use live models running continuously. It remains to be seen how these systems perform when they have to react to real-time changes constantly.
Another open question involves the reproducibility of these results over multiple independent development cycles. While the current study shows high success rates, researchers need to determine if these outcomes are consistent every time a new program is built. This information will be vital for companies looking to deploy these technologies in real-world settings.
What to watch
- How these systems handle objects that change shape significantly.
- Comparisons between pre-set programs and live, continuous AI models.
Try it
Look up the ACE-CaP-0 method to see how researchers are structuring these robot programs.
Sources and links
- ReportingBornCity: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by BornCity · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
