Democratizing Robot-Use Agents: A Unified Harness with Open Models

TL;DR

uni-robot-agent (Yuan, Zuo, Huang, Zhang, Park et al.; Berkeley + Tsinghua) is a unified harness for competitive open-weight Robot-Use Agents (RUAs) that also supports further training — the explicit democratization bet is that open-weight models can match frontier proprietary RUAs if the harness contract is right. Without privileged inputs, an off-the-shelf Qwen3.8-Flash-Next driven by the harness reaches RoboDojo Score 31.39 on the shared RoboDojo RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies substrate, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. The paper argues effective robot use ties to four general capabilities, suggesting future RUA training should develop them jointly rather than only the harness slot.

Key claims

Current RUA research depends heavily on frontier proprietary models (Astra, Fable, Opus), which limits further training and reproducibility; uni-robot-agent closes this gap with a unified harness that supports both inference-time deployment and training-time use of open-weight models [Abstract].
With no privileged inputs, an out-of-the-box Qwen3.8-Flash-Next under the unified harness reaches RoboDojo Score 31.39, beating the published GPT-6 Astra-RoboProbe result at 28.97 [Abstract].
The same Qwen + harness configuration beats “many VLA/WAM baselines” on RoboDojo and performs competitively on 19 real-world tasks [Abstract].
Analysis across model generations, families, and sizes links effective robot use to four general capabilities (named but not surfaced in the abstract) — argues these should be developed jointly in future RUA training rather than offloaded to the harness [Abstract].

Method

The arxiv landing page carries only an abstract; method details below reflect what the abstract commits to and are not a reconstruction of the full paper. uni-robot-agent is a harness — the LLM-as-controller row of the recipe-lever board — exposing some fixed interface to a frontier open-weight VLM so the VLM can perceive, plan, and act on a robot in both simulation (RoboDojo) and the real world without additional policy training. Two key framing moves per the abstract: (1) it does not rely on privileged inputs (no ground-truth state leakage into the agent’s observation), and (2) it is training-ready — the same harness that drives inference can be swapped under a model being fine-tuned, so future work can co-adapt the model and the harness contract.

Evaluation uses RoboDojo RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies as the simulation substrate (42 tasks across five capability dimensions on Isaac Sim, plus 18 real-world tasks on standardized remote-cloud hardware) and 19 real-world tasks of the authors’ own definition. The flagship model configuration is Qwen3.8-Flash-Next without finetuning. The comparison target is the published GPT-6 Astra-RoboProbe result on RoboDojo (28.97) plus VLA/WAM baselines on the same substrate.

Results

Headline numbers reported in the abstract:

RoboDojo Score 31.39 for Qwen3.8-Flash-Next under uni-robot-agent, vs 28.97 for the published GPT-6 Astra-RoboProbe result — the sharpest filed demonstration that an open-weight model plus the right harness can outperform a frontier proprietary RUA on the same benchmark [Abstract].
Beats “many VLA/WAM baselines” on RoboDojo [Abstract]. For context on the same substrate: Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement COAP reports 70.24% on RoboDojo’s 42 bimanual tasks with no model at test time, and InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data InternW0-Δ reports 23.9% as the strongest prior-filed WAM on the same bench. Where 31.39 sits in that scoring convention is not resolvable from the abstract.
Competitive performance across 19 real-world tasks (no per-task numbers in the abstract) [Abstract].
Four general capabilities named as the common thread across model generations / families / sizes that effective robot use depends on; the four are listed in the paper body but not surfaced in the abstract [Abstract].

Why it’s interesting

This paper lands directly on the open vs. frontier axis of the LLM-as-controller row of the VLA Models recipe-lever board — the same cell occupied by OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA (zero-abstraction ROS 2 CLI + Claude Code), RobotUse: Allocating Computation, Context, and Decisions RobotUse (three-role Main/Subagent/Backend), Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy Astranomy (GPT-6 Astra with persistent Codex sessions), and ARC: A Reasoning Recipe for Robot Foundation Models ARC (reasoning-wrapper over π0.5/Cosmos3-Nano). It makes two load-bearing moves none of them make jointly: it fixes the open-weight model (Qwen3.8-Flash-Next) in the controller slot and claims a specific head-to-head win over the frontier Astra entry on the shared RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies RoboDojo substrate — Astranomy’s main cost-vs-capability finding was that Astra at $36.40 and ~30.9M tokens per successful episode delivers 74.8% on RoboLab-120 while open models in a Direct-Policy configuration collapse (Sol 33.3%, Luna 2.0%). uni-robot-agent challenges the implicit “frontier model + any harness beats open model + any harness” assumption that reading left standing, and reframes the design surface explicitly as a model × harness system. The second move — that the harness also supports further training — directly connects this cell to WHALE — joint harness + weights optimization via alternating block ascent WHALE’s alternating-block-ascent recipe for jointly optimizing model and harness, and to the broader Agentic Software Engineering substrate where “coding agent + documented API over native robot stack” is what OpenRUA: Robot-Use Agents Are Zero-Shot Visuomotor Policies OpenRUA argues is actually transferring. The headline Qwen-beats-Astra claim should be read against two caveats surfaced elsewhere on the board: (a) Astranomy: An Empirical Study of GPT-6 Astra as a Native Robot Policy Astranomy’s finding that persistent conversation history was the dominant lever for Astra, so the comparison with “published Astra-RoboProbe” depends on what harness Astra ran under; (b) the four-capability analysis is the paper’s own framing of the gap, and whether jointly training those capabilities actually closes it remains to be seen on an open-weight base.