OpenVLA 2 released with improved generalisation - RobotWale News

OpenVLA 2 released with improved generalisation

The OpenVLA consortium has officially announced the release of OpenVLA 2, a major update to its open-weight Vision-Language-Action model designed for humanoid robotics. The release, made public earlier this week, marks a significant leap in the model's ability to generalise across diverse robotic platforms and environments. Developed by a collaboration of leading AI researchers and robotics engineers, the new iteration aims to solve the "sim-to-real" gap that has long hindered the deployment of autonomous robots in unstructured settings.

OpenVLA 2 introduces enhanced transformer architectures that allow the robot to interpret visual inputs and execute physical actions with greater precision. According to the release notes, the model demonstrates a 30% improvement in task generalisation compared to its predecessor, OpenVLA 1.0. This improvement is particularly critical for humanoid robots tasked with complex manipulation tasks, such as assembly line work or domestic assistance, where environmental variables often change unexpectedly.

Key technical advancements in the new release include:

• Adaptive Reasoning: A new module that allows the robot to adjust its action policy based on real-time visual feedback.
• Multi-Robot Coordination: Support for collaborative tasks where multiple robots share a single language model.
• Low-Latency Inference: Optimised for edge computing, enabling faster decision-making on local hardware without heavy cloud dependency.

For the Indian robotics ecosystem, the release of OpenVLA 2 offers a timely opportunity for cost-effective automation. The OpenVLA team has confirmed that the model weights will remain open-source, allowing Indian startups to fine-tune the model on local datasets without incurring licensing fees. However, commercial support packages for enterprise integration are being priced competit