Harness VLA Redefines Embodied Intelligence Execution and Beats NVIDIA Cap‑X

System Overview

Harness VLA introduces a Harness Layer on top of a frozen Vision‑Language‑Action (VLA) model. The frozen VLA is wrapped as a VLA_ACT Primitive that handles contact‑dense actions such as grasp, place, button press, and drawer/door manipulation. Analytic Primitives ( MOVE_TO, ROTATE_WRIST, SET_GRIPPER, RELEASE) perform stable spatial motions. An Agentic Planner learns, during an exploratory bootstrapping phase, when to invoke each primitive, how to stage the scene, and how to recover from failures.

Two‑Phase Operation

During the Exploratory Bootstrapping Phase the planner interacts with the environment, records successful primitive sequences as Task‑Specific Memory, and aggregates cross‑task failure rules into Global Memory. In the Deployment Evaluation Phase the planner uses these memories to re‑bind targets, adjust the robot state, and decide whether to call the VLA_ACT Primitive or Analytic Primitives.

Key Design Insights

Failures often arise from acting on the wrong target, state, or task stage; re‑binding targets and adjusting states before VLA calls mitigates this.

Calling the frozen VLA only when the robot is in a suitable state dramatically improves robustness; after each call the planner decides to continue, adjust, or retry.

Analytic Primitives off‑load non‑contact execution, allowing the frozen VLA to focus on contact‑dense skills while the planner handles navigation, positioning, and recovery.

Evaluation Benchmarks

LIBERO‑Pro (target re‑binding and layout perturbations). Baseline RLinf pi05 model success 50.0 %; Harness VLA (Claude Code planner) achieves 82.4 %, surpassing Pi_RLinf (50 %), NVIDIA Cap‑X (18.2 %) and Berkeley RATS (43.8 %).

RoboCasa365 (real‑world kitchen tasks). Using the frozen RLDX‑1 baseline rates 60 %, 21.3 %, 5 % for ATOMIC‑SEEN, COMPOSITE‑SEEN, COMPOSITE‑UNSEEN; Harness VLA raises them to 91.6 %, 56.3 %, 15 % respectively.

RoboTwin C2R (zero‑shot transfer from Clean to Randomized dual‑arm scenes). Harness VLA reaches 58.4 % success, higher than Cap‑X (47.9 %), GR00T‑N1.7 (20.7 %) and StarVLA (10.6 %).

Few‑Shot vs Zero‑Shot

In few‑shot evaluation the planner first discovers a successful primitive trajectory for each task (Task‑Specific Memory) on Seed 0, then reuses it on other seeds. In zero‑shot evaluation the planner does not retrieve memory and relies solely on online observations; performance drops (e.g., LIBERO‑Pro Goal 31.0 % vs 79.0 % under task‑T) but still exceeds Cap‑X, showing that the planner can still organize VLA calls and Analytic Primitives without memory.

Memory‑Guided Call Scheduling

Success‑rate curves show that allowing only a few VLA_ACT calls per episode yields most of the performance gain, indicating that sparse, targeted VLA invocations combined with analytic pre‑ and post‑processing are sufficient.

Open‑Source Release

Code and documentation are released as part of the RPent (Recursive Physical Agent) ecosystem: https://github.com/RLinf/RPent. Project page: https://harnessvla.github.io/. ArXiv paper: https://arxiv.org/pdf/2607.08448.

Illustrative Figures