Physical Intelligence Announces “Knowledge Insulation,” a New VLA Training Method; π0.5+KI Aims to Combine Fast Training and Generalization

Physical Intelligence announced a new method for VLAs, “knowledge insulating VLAs,” on May 28, 2025. At its core, the design adds an action expert for continuous action outputs without allowing gradients from the expert to flow back to the VLM backbone. The company calls the method π0.5 + KnowledgeInsulation (KI). It presented an architecture in which the backbone is trained with FAST-tokenized discretized actions, while the action expert handles continuous actions. The paper compares this approach with tokenized discrete-action methods such as RT-2 and OpenVLA, and with the continuous-output approach used by the π0 series.

If Physical Intelligence’s claims are consistent with its primary materials, VLA performance differences depend not only on model size but also on how knowledge in the VLM backbone is protected during training. Separating the action expert for robot control from the VLM backbone for language and visual representations offers a basis for considering how to combine real-world accuracy with training efficiency. The company also said it uses general VLM data and robot planning data, underscoring that robot training does not rely on a single dataset or a single control head.

The novelty is not just adding continuous action outputs to a robot-focused VLA, but explicitly addressing how to preserve the VLM backbone’s knowledge as part of the training design. As an extension of the existing π0 series, the overall arrangement—using an action expert for continuous control—remains the same. But the company says it aims to make web-scale pretrained representations less vulnerable to disruption by changing how gradients flow. Our related article, 36Kr Reports Interview With Wang Xingxing of Unitree Robotics, Discusses Breakthroughs in Physical AI, examined where the breakthrough in Physical AI might come from. This announcement addresses that question at a deeper technical level: how to design a system that preserves internal model knowledge while enabling robot control. Put another way, the issue is not only the accuracy of a robot’s actions, but also how much of the VLM’s language-understanding representations can be maintained during training. From an industry perspective, the focus shifts from hardware alone to the combination of a VLM backbone, an action expert, continuous-control training methods and robot planning data. The announcement again highlights a trade-off in VLA design: tokenized discrete actions are easier to train but slow at inference, while continuous outputs are smooth but have more difficult training dynamics. Questions to watch include how well the KI design generalizes across real-world tasks and under what conditions it is effective compared with approaches such as freezing the backbone or blocking gradients.

According to the published materials, π0.5 + KI consists of two components: a VLM backbone and an action expert. The backbone learns representations using FAST-tokenized discretized actions, while the action expert generates continuous actions through flow matching; its gradients do not propagate to the VLM backbone. The materials also say that with conventional action expert designs, the VLM backbone’s representations may be exposed to unfavorable training dynamics, potentially weakening adherence to language instructions and knowledge acquired through web-scale pretraining. The company says π0.5 + KI also uses general VLM data and robot planning data for additional training.

For Japanese robotics companies and research institutions, how VLM backbones, action experts and robot planning data are combined could become a competitive factor, in addition to the control of the hardware itself. In particular, a design that trains continuous control while maintaining adherence to language instructions could connect with Japan’s strengths in real-world testing and training-data design for industrial applications. However, the primary materials do not mention any connection with Japanese companies or identify potential applications.