Open-weight robot agent outperforms proprietary models on zero-shot manipulation benchmark

Researchers have released a unified testing harness that lets open-weight models run robot-use agents without proprietary access. The project, called uni-robot-agent, reports that an out-of-the-box Qwen3.8-Flash-Next model scored 31.39 on the RoboDojo benchmark, above the published GPT-6 Astra-RoboProbe result of 28.97. The work targets a practical bottleneck: robot-use agent research has depended on frontier proprietary models, which limits anyone who wants to train and compare their own systems.

Robot-use agents (RUAs) take instructions and act in physical environments, a task that has drawn interest for zero-shot manipulation - performing a task without prior training on it. The team says their harness supports further training, not just inference, and requires no privileged inputs. Alongside the benchmark score, the open model performed competitively on 19 real-world tasks and beat several vision-language-action (VLA) and world-action-model (WAM) baselines.

What the harness changes for researchers

Much of the published RUA work relies on closed models, so outside labs can't inspect, fine-tune, or reproduce results on the same footing. A harness built around open weights lowers that barrier. Teams can run the same evaluation, train on top of the base model, and compare results without negotiating API access. For groups working on robotics and agent systems, that changes the cost and speed of iteration - a consideration that also runs through broader AI R&D Engineering Courses aimed at research and development engineers.

Four capabilities behind effective robot use

The researchers analyzed results across model generations, families, and sizes. They link effective robot use to four general capabilities, and argue that future RUA training should develop them jointly rather than in isolation. The abstract does not name the four in detail; the finding is framed as a direction for training design. The core claim: capability in one area doesn't transfer cleanly on its own, so training pipelines should target them together.

The comparison figures come from RoboDojo, a benchmark used to score agent performance on manipulation tasks. The 31.39 result was produced without privileged inputs, meaning the model worked from the same information a real deployment would have. The 19 real-world tasks add a check beyond simulation, though the abstract reports the outcome as competitive rather than a specific score.

Why this matters for robotics and AI researchers

If open-weight models can match or beat proprietary ones on standardized robot benchmarks, the entry cost for RUA research drops. Teams can train, test, and publish against a shared harness instead of building evaluations from scratch or paying for frontier model access. The open question is training: the paper's argument that four capabilities should be developed jointly gives labs a concrete target for the next round of work. Researchers planning that work can find related material in AI Agent Courses covering agent design and automation.