Robot agents using open models can beat a published proprietary baseline

The authors report that an off-the-shelf open-weight model surpassed a published RoboDojo result from a proprietary model, without privileged inputs, and performed competitively on 19 physical manipulation tasks.

Top University

Jiarui Yuan · Yuxin Zuo · Xiaoyu Huang · Wenbo Zhang · Kwanyoung Park · Junlin Yang · +22 more

Tsinghua University · UC Berkeley

Research Digest·3 November 2026·2 min read

Illustration: Robot agents using open models can beat a published proprietary baseline

In one line

An open-weight agent, using a unified harness without privileged inputs, surpasses a proprietary model on a robot manipulation benchmark.

Before you read this

Three ideas this paper builds on

agent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.161 papers →fine-grained spatial reasoningReasoning precisely about object positions, orientations, and spatial relations across planar, depth, and temporal dimensions for robotics and navigation.60 papers →code as actionA paradigm where actions are expressed as executable code, such as method calls or program synthesis, rather than direct manipulation.27 papers →

Read it as

Robot agents that can manipulate objects are usually powered by expensive, proprietary models. This makes it hard for researchers to study and improve them.

A new system called uni-robot-agent provides a standard way to connect open-weight models (models whose internal parameters are publicly available) to robots. It includes reliable tools and efficient execution, so a model like Qwen3.8-Flash-Next can be used out of the box.

On the RoboDojo benchmark, this open-weight agent scored 31.39, beating the proprietary GPT-6 Astra system's score of 28.97. This shows open models can be competitive for robot use, enabling wider research and local deployment.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

robot-use agent agent harness fine-grained spatial reasoning code as action privileged information long-horizon robot manipulation