NVIDIA showcases research at CoRL on robots learning precision tasks in simulation

In an official blog post published Nov. 9, 2026, NVIDIA said it presented research with partners at CoRL on robots practicing precision tasks in simulation. Examples included gear alignment, cable connections and small-part assembly. The blog said the use of simulation in robot learning is expanding from locomotion tasks such as walking, running and jumping to tasks involving fine manipulation.

NVIDIA's blog showed that simulation for robot learning is being used not only for locomotion such as walking, but also to practice tasks such as gear alignment and parts assembly. When such tasks are considered for manufacturing processes, reproducibility and precision on physical robots are conditions for adoption. The blog did not provide performance figures or results from testing on physical robots.

The key point is that simulation is being used not only for locomotion, but also for manipulation tasks such as gear alignment and cable connections. The blog described robots as “practicing” tasks in simulation, but did not present a method for transferring the learning to physical robots or report success rates in real-world settings. What can be gleaned at this stage is a change in the tasks being learned, not that automation of real-world work has been achieved. In related coverage, NVIDIA advances Physical AI in South Korea across semiconductors, robotics and autonomous driving, published Nov. 9, 2026, examined industrial applications of Physical AI based on manufacturing and engineering. While that article focused on the breadth of industrial fields, including semiconductor manufacturing, robotics and autonomous driving, this blog focuses on the granularity of tasks performed by robots. Even taken together, the two reports do not establish any connection between specific robots or deployment plans. But when assessing industrial applications, the tasks robots are trained to perform in simulation are a consideration alongside computing infrastructure and fields of application. An industry issue is how to extend learning environments centered on locomotion to manipulation tasks involving contact and positioning. While locomotion primarily involves trajectories and posture, parts assembly also involves the relative positions of objects and behavior during contact. However, the blog did not specify the simulation methods or how force and tactile information were handled, so it is not possible to determine which elements were modeled. Questions to watch include how well results from simulation-based learning can be reproduced on physical robots, which metrics are used to evaluate task accuracy and completion time, and how well the systems can handle differences in objects and environments. The information provided does not identify the participating partners or robot models, or specify the training data, performance figures or whether physical-robot testing was conducted. Without those details, it is not possible to assess the potential for application in mass-production processes beyond the scope of the research.

NVIDIA said simulation has been used as a test environment for teaching robots to walk, run and jump. The manipulation tasks presented this time were gear alignment, cable connections and small-part assembly. The blog did not name the participating partners or specify robot models, simulation methods, performance figures or results on physical robots.