Panasonic Holdings Announces 'Proxy3D' to Strengthen VLMs' 3D Capabilities at CVPR 2026

Nikkei xTECH reported on Sept. 10, 2026, that Panasonic Holdings announced Proxy3D, a new technology to enhance VLMs' spatial awareness, at CVPR 2026, held in June 2026.

The research Panasonic Holdings presented at CVPR 2026 with Tsinghua University and UCB addresses a weakness in VLMs: understanding physical space. Robots and AI systems operating in real-world environments need to handle not only language and images but also 3D spatial relationships, making the technology's achievable accuracy and computing requirements key questions.

What is new here is that Panasonic Holdings presented Proxy3D, which targets the relatively specific weakness of 3D recognition in VLMs, and announced it at the international conference CVPR 2026 as joint research with Tsinghua University and UCB. Rather than improving the capabilities of foundation models in general, the work focuses on understanding positions and distances in the real world. The issue therefore appears to concern robotics and real-world understanding, where spatial awareness is needed, rather than generative AI in general. In related articles, Thinker director speaks at AI robotics study group, presents solutions to contact challenges covered contact challenges for robotic hands, while Panasonic Holdings and SHIN-JIGEN joint research wins Outstanding Presentation Award at robotics conference reported recognition for the company's robotics-oriented joint research. Proxy3D can be read as taking that research further toward spatial understanding, though it is difficult to determine without consulting the original source whether it extends the contact control and joint research discussed in those articles or addresses a different issue closer to VLM preprocessing or perception. From an industry perspective, the focus shifts from the amount of model training itself to how 3D space is represented and how many fewer tokens are needed to process it. This could affect how information is compressed for robot control and on-site perception, with potential implications for computing resources, training data preparation and the sequence of implementation in physical systems. However, the excerpt does not provide Proxy3D's performance metrics, target tasks or implementation form. Questions to follow include which VLMs it connects to, what evaluation criteria are used for 3D recognition, and whether it is ready for use in real robots or on-site systems.

For Japanese companies, a key issue is how to integrate 3D spatial awareness into AI for robotics. In particular, whether VLMs' spatial understanding can be linked to physical-system challenges such as contact and alignment could affect the competitiveness of Japan's manufacturing and robotics technologies.