HarnessVLA lifts LIBERO-Pro success to 82.4% by freezing the VLA and adding a planner

机器之心 · wechat · 2026-07-22

Tsinghua and partners release HarnessVLA, a new embodied AI framework that keeps the VLA frozen and adds an agentic planner plus a harness layer to coordinate calls, execution, and recovery.

In LIBERO-Pro disturbance tests, it raises success from 50% to 82.4%, beating PiRLinf, NVIDIA Cap-X, and Berkeley RATS. The framework is designed to be model-agnostic and can also work with WAM and other embodied foundation models.

Key idea: many robot failures are not because the VLA cannot act, but because it acts on the wrong target, in the wrong state, or at the wrong task stage. HarnessVLA learns how to organize and reuse existing skills more reliably rather than retraining the base model.

Original post →

More from Embodied

Embodied channel →