Our industry has made significant progress in recent years to enable industrial robots to better perceive their surroundings. This ability to understand, interpret, and act on various information is crucial for robots to operate in complex real-world environments. While the perception team at Intrinsic has made great headway in overcoming current limitations through AI-enabled perception solutions, we were also proud to help push the boundaries of what’s possible for industrial applications by sponsoring OpenCV’s 2025 Perception Challenge for Bin-picking. Congratulations to the winning teams who were celebrated at the 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) and to the more than 450 participating teams who worked to develop robust 6 Degrees-of-Freedom (6DoF) pose estimation solutions. 

‍

Looking back at the field of computer vision, the potential for computers to “see” hit a pivotal moment when transitioning from a traditional vision approach to a deep learning model. This may now seem obvious, especially as self-driving cars have increased awareness of how AI-enabled perception can work seamlessly with real-world hardware to deliver new value, but for a long time our industry faced many unsolved problems. Today, we’re advancing at great speed and scale, with leading methods quickly leapfrogged and innovation only accelerating.

‍

In industrial robotics, the opportunities for AI-enabled and specialized perception capabilities are especially vast when it comes to unlocking new value for solutions builders and manufacturers. Last year at SIGGRAPH Asia, the perception team at Intrinsic unveiled a new plenoptic 3D vision system designed specifically for industrial robotics applications. This novel multi-camera multimodal vision system simultaneously improves robot efficiency and precision while also making it more feasible for the system to generalize across different cameras, lighting, and materials in industrial settings, from factories to fabrication shops. We’ve since integrated the foundation models directly into Intrinsic Flowstate, our digital twin developer environment.

‍

Still, a key challenge our industry faces in pose estimation metrics is the disconnect between how vision algorithms are evaluated and how they are ultimately used. Using Intrinsic open sourced datasets and a real robot cell, our own evaluations are based on the direct impact of whether a robot can successfully complete a pick-and-place task in common industrial settings. Throughout the bin-picking challenge, we made real hardware remotely usable for the participants through the Intrinsic platform, which resulted in an evidence-based approach as they tested the real-world performance of new algorithms:

‍

• The results of the bin-picking challenge further demonstrate the efficacy of the multiview perception framework. Most of the top-performing solutions utilized multiview strategies for object pose estimation, serving as a strong validation of our proposed methodology for robust perception in industrial automation. Being able to interpret diverse objects more intelligently can also enable even more seamless integration into other workflows.

‍

• The prize for a one-shot solution stood out for special recognition, as AI radically reshapes object identification despite highly limited training data. Advancing open source datasets and metrics through industry collaborations encourages our broader ecosystem to further accelerate innovation and the adoption of new automation solutions.

‍

What inspired us throughout the Perception Challenge for Bin-picking was watching the winning teams redefine what’s possible in robotic automation — right in front of us, in real-time. Critically, we confirmed the ease of integrating their diverse, third-party solutions directly into our Flowstate workflow, successfully running them on our hardware for precise pick-and-place tasks. More than a technical achievement, this demonstrates how building with an AI-integrated platform like Intrinsic, can unlock new value through next-gen robotic skills deployed in real-world systems.

As we work to overcome some of the more persistent challenges we continue to face around lighting conditions, the translucency and geometries of certain materials, and the workpieces involved, these leaps in perception will further enhance precision, enable reusable workflows, and improve scalability. The bin-picking challenge helped show what’s possible for our industry. We’re grateful for our partnership with OpenCV in making this event a success, and again congratulate all the teams whose ingenuity is helping to solve critical manufacturing problems.