OpenTau Open-Source VLA Training Toolchain Announced at CES
At CES 2026, Tensor announced the official open-source release of OpenTau (τ), a powerful AI training toolchain that accelerates the development of Vision-Language-Action (VLA) foundation models. These models form a critical building block for the next generation of Physical AI systems.
At Tensor, we push the frontier of large foundation models for Physical AI. VLA models combine vision, language, and action in a single multimodal foundation model. This design helps intelligent systems understand the world, reason about it, and act within it. The field increasingly recognises this approach as a leading paradigm for Embodied AI, with applications across autonomous driving, robotic manipulation, and navigation.
OpenTau (τ) is Tensor’s open-source AI training platform for frontier VLA models. The platform makes large-scale AI training reproducible, accessible, and scalable. By releasing OpenTau to the global research and development community, Tensor makes advanced training capabilities available beyond closed, proprietary environments, helping accelerate innovation across the industry. We believe the most meaningful technological advances depend on clarity, rigour, and reproducibility. Open-sourcing supports that philosophy by providing scientific transparency and enabling independent validation.
"At Tensor, we believe meaningful progress in Physical AI requires transparency," said Jay Xiao, Founder and CEO of Tensor. "OpenTau is our way of giving back to the research and developer community that has helped advance this field. By open-sourcing our training toolchain, we're supporting broader collaboration—so everyone can build, experiment, and move faster together."
OpenTau brings state-of-the-art VLA training capabilities to the open-source community. It supports co-training on an adjustable mixture of heterogeneous datasets. It uses discrete-action modelling to accelerate convergence of the Vision-Language Model (VLM). It adds knowledge insulation between the VLM backbone and the action expert. It applies VLM dropout techniques to reduce overfitting. It includes a reinforcement learning pipeline built specifically for Vision-Language-Action models. It also includes more capabilities that help teams train and evaluate VLA foundation models with greater control and reliability.
OpenTau prioritises reproducibility and extensibility. The platform lets researchers and developers test advanced training strategies that often remain difficult to implement or inaccessible outside large-scale industrial research settings. We believe open research strengthens the whole AI industry over time.
Tensor invites researchers, developers, and builders to explore OpenTau, contribute to its evolution, and help shape the future of Physical AI. You can support the project by starring the repository, forking and experimenting with the codebase, and building or extending frontier Vision-Language-Action models.