Embodied AI May Enter the Token Era: Tsinghua Team Proposes Harness VLA, Ushering in a Paradigm Shift

Today, the team led by Professor Yu Chao at Tsinghua University, together with Zhengxing Innovation and Infinigence AI, officially released and open-sourced Harness VLA. This technical achievement introduces the Harness Layer, widely used in digital intelligence, into embodied AI systems for the first time. While keeping the VLA model frozen, it uses an Agentic Planner to uniformly organize model invocation, task execution, and failure recovery, moving robots from "relying on a single end-to-end model" toward a new architecture in which "model capabilities and system capabilities work in concert," significantly improving robot generalization in complex real-world environments.

In the more challenging LIBERO-Pro perturbation benchmark, Harness VLA raises the success rate to 82.4%, significantly exceeding Pi_RLinf (50%), NVIDIA Cap-X (18.2%), and Berkeley RATS (43.8%), validating the feasibility of the new-generation robot intelligence architecture of "model + system." More importantly, Harness is not designed for a specific model but is a general framework oriented toward the system layer. In addition to VLA, it can also be combined with embodied foundation models such as WAM, unleashing the capabilities of different foundation models through a unified task execution layer and providing a more general organizational approach for future embodied AI systems.

Over the past two years, Vision-Language-Action (VLA) models have brought success rates on standard benchmarks for robotic manipulation to near saturation.

For example, the Pi 0.5 model trained on 130 LIBERO tasks achieves an average success rate as high as 96%. From grasping and transporting to organizing objects, most tasks can already be completed reliably.

But these impressive numbers do not really answer a more practical question: can robots still work properly after leaving standard test environments?