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

Abstract: Robot-use Agents (RUAs) have shown promising zero-shot manipulation capabilities, but reliance on frontier, proprietary models limits research into further training. To democratize RUA research, we introduce uni-robot-agent, a unified harness for competitive open-weight robot use that supports further training. Without privileged inputs, an out-of-the-box Qwen3.8-Flash-Next achieves a RoboDojo Score of 31.39, surpassing the published GPT-6 Astra-RoboProbe result of 28.97 and many VLA/WAM baselines, and performs competitively on 19 real-world tasks. Our analysis across model generations, families, and sizes links effective robot use to four general capabilities, suggesting that future RUA training should develop them jointly.

• The paper defines `uni-robot-agent`, a unified interaction harness that enhances reproducible RUA research with open-weight foundation models.
• Researchers evaluated models achieving up to 25.00% success rate without privileged simulator state, outperforming published baselines in simulation with a RoboDojo score of 31.39 for Qwen3.8-Flash-Next.
• Qwen3.8-Flash-Next exhibitions viable physical manipulation tasks, scoring 49.50 in real-world tests, though the unstructured environment leads to certain failure cases across various task profiles, verifying embodiment skills are significantly attributed to respective, not uniform, task competence.

Research problem and central claim

“Democratizing Robot-Use Agents: A Unified Harness with Open Models” presents uni-robot-agent, a modular interaction harness intended to make robot-use-agent (RUA) research reproducible, extensible, and compatible with open-weight foundation models (2611.01560). The paper addresses a systems-level bottleneck in RUA research: existing demonstrations of zero-shot robotic manipulation depend heavily on proprietary frontier models, leaving researchers unable to inspect, fine-tune, cache, or otherwise modify the underlying policy.

The paper’s central claim is that a sufficiently structured harness can expose substantial robot-use capability in open-weight models without privileged simulator state, task-verifier feedback, or distillation from closed models. The strongest reported result is obtained by Qwen3.8-Flash-Next, which achieves a RoboDojo score of 31.39 and a success rate of 25.00% over 252 episodes. This exceeds the published GPT-6 Astra–RoboProbe result of 28.97, although the comparison must be interpreted carefully: GPT-6 Astra reaches 51.90 when evaluated with uni-robot-agent itself, demonstrating that the harness and model jointly determine performance. The paper therefore supports a narrower but important conclusion: an open-weight model using the proposed harness exceeds a previously published closed-model baseline using a different harness.

The work combines four contributions. First, it defines a unified interface spanning models, environments, robot embodiments, tools, and feedback. Second, it introduces execution mechanisms that overlap model inference with physical action while preserving resource ordering. Third, it evaluates open-weight models under non-privileged observation conditions in simulation and on real robots. Fourth, it analyzes robot use as an emergent combination of visual understanding, spatial grounding, geometric coding, and embodied reasoning rather than as an isolated robotics-specific capability.

Figure 1: System overview of uni-robot-agent, combining shared tools, environment adapters, interchangeable runtimes, and simulation/real-world evaluation.

Harness architecture and interaction model

The system is organized into five logical layers: agent runtime, orchestration, shared tools, domain services, and concrete backends. The agent layer manages model interaction and context; orchestration schedules requests and tool execution; shared tools expose common robot operations; domain services manage perception, environment, and action-model resources; and backends implement simulator, robot, and model-specific functionality. This separation allows a runtime, perception service, or robot backend to be replaced without rewriting the remaining layers.

Figure 2: Five logical layers separating model interaction, scheduling, shared tools, services, and environment-specific backends.

The unified interface abstracts motion, perception, skills, and episode control behind common arguments, units, and feedback formats. Environment adapters translate these operations into native simulator or robot APIs. This avoids implementing each tool separately for every environment: $N$ shared tools can operate across $M$ adapters rather than requiring $N \times M$ tool-environment implementations. More importantly, the common interaction schema makes trajectories collected across simulation and physical systems structurally compatible with later training.

The harness also treats physical consistency as an explicit systems requirement. Images, camera calibration, and robot state are bound to immutable snapshots, preventing localization from combining an image captured at one robot pose with calibration or state from another. Request identities and execution states prevent communication failures from causing repeated physical actions. Motion and gripper control can be issued jointly, while monitored motion advances through measured intermediate states and reports residual errors, planning rejections, stalls, divergence, and stopping reasons.

These mechanisms address failure modes that are not reducible to model intelligence. A LLM may issue a valid pose command, but reaching that pose does not imply that an object was grasped or placed successfully. By returning numerical pose errors together with visual outcomes, the harness gives the model evidence for replanning. The design therefore treats tool feedback as part of the policy’s observation space rather than as an implementation detail.

Figure 3: Paired design choices for state consistency, execution reliability, coordinated control, monitored motion, diagnostic feedback, and context management.

Efficient execution through resource-aware pipelining

A key architectural contribution is the two-stage pipeline between model reasoning and tool execution. The model can submit a background tool call and continue reasoning while the robot executes it. Calls sharing an exclusive resource remain ordered, whereas independent calls can run concurrently. A subsequent decision that depends on a tool result must still wait for that result, and queued actions can be cancelled with flush_queue when new feedback invalidates the plan.

This is a more constrained form of concurrency than unrestricted parallel execution. The system overlaps computation and action without violating arm-specific ordering or allowing dependent actions to execute against stale state. The distinction is important for physical systems, where naïve parallelism can produce collisions, contradictory commands, or plans based on observations that no longer correspond to the robot configuration.

The empirical speed results are substantial but based on a limited evaluation. In five selected tasks, pipelining reduced episode duration by 4.8% to 73.7% relative to serial execution. The largest example, swap_T, decreased from 34.7 minutes to 12.5 minutes; press_by_number decreased from 25.0 to 13.7 minutes. The proportion of physical execution time overlapping with model requests ranged from 84.5% to 93.8%, compared with zero overlap under serial execution. These results imply that inference latency need not be fully additive with physical execution time, provided that dependency and resource constraints are represented explicitly.

The qualitative trajectories show why overlap is useful but not sufficient. During transport or release, the model can prepare geometric calculations, clearance checks, or orientation candidates. However, execution feedback can invalidate the prepared plan, requiring replanning. Pipelining therefore improves throughput while preserving a feedback-controlled execution loop; it does not eliminate the need for state verification.

Figure 4: Model requests overlap most physical execution time under the pipelined scheduler, whereas serial execution has no overlap.

Figure 5: Pipelining reduces episode duration across five tasks, with reductions ranging from 4.8% to 73.7%.

Simulation performance with open-weight models

The simulation evaluation uses 42 RoboDojo tasks covering generalization, precision, long-horizon execution, memory, and open-vocabulary instruction following (Chen et al., 5 Jul 2026). Each primary configuration is evaluated with three seeds and two layouts, producing 252 episodes. The principal experiments exclude simulator depth and task-verifier feedback.

Model and harness	Success rate	RoboDojo score
GPT-6 Astra + RoboProbe	22.48%	28.97
DeepSeek-V4.1-Flash + uni-robot-agent	21.39%	27.11
Qwen3.8-Flash-Next + uni-robot-agent	25.00%	31.39
GPT-6 Astra + uni-robot-agent	46.67%	51.90
Qwen3.8-Flash-Next + GAP (Hybrid)	16.11%	21.01

Qwen3.8-Flash-Next with the proposed harness surpasses the published Astra–RoboProbe result despite using an open-weight model and no privileged inputs. DeepSeek-V4.1-Flash also exceeds the listed VLA and WAM baselines, with a score of 27.11. The harness comparison is particularly informative: replacing uni-robot-agent with GAP reduces Qwen’s success rate by 8.89 percentage points and its score by 10.38 points; RoboProbe and RPent perform substantially worse still. This result attributes a meaningful portion of performance to interaction design, feedback representation, scheduling, and tool semantics rather than to the model alone.

The comparison also exposes an important qualification. GPT-6 Astra obtains 51.90 with uni-robot-agent, far above Qwen’s 31.39. Thus, the paper does not establish that open-weight models generally outperform closed models under matched conditions. It establishes that the proposed harness allows at least one open-weight model to exceed a previously published closed-model configuration and to achieve competitive performance under realistic information constraints.

Figure 6: Robot-use performance varies across model generations, families, and sizes; both model capability and training history affect RoboDojo outcomes.

Real-world manipulation

The physical evaluation uses 19 tabletop manipulation entries on a dual-arm platform with three Intel RealSense D405 cameras. Tasks are grouped into spatial understanding, object generalization, instruction following, compositional generalization, and unseen tasks. Scores are rubric-based and allow partial credit, so the aggregate is not equivalent to full-episode success.

Qwen3.8-Flash-Next obtains a total score of 49.50, matching the 47.75 reported for fine-tuned $\pi_{0.5}$ and exceeding Dream-1’s 35.50. GPT-6 Astra reaches 75.50. Qwen performs particularly well in compositional generalization, scoring 13.00 out of 20, and matches Astra on the unseen-task subtotal at 8.00. It is weaker in object generalization, scoring 6.00 compared with 12.50 for $\pi_{0.5}$ and 12.50 for Astra.

The real-world comparison is not fully controlled with respect to sensory inputs. uni-robot-agent configurations receive depth from the RealSense cameras, whereas $\pi_{0.5}$ and Dream-1 consume RGB inputs. In addition, $\pi_{0.5}$ is fine-tuned on 10 hours of demonstrations per task, while the RUA configurations are evaluated zero-shot. The results therefore show competitive zero-shot physical manipulation, but they do not constitute a strict modality-matched comparison between the systems.

The task-level results illustrate the different generalization profiles. Qwen succeeds on color-conditioned cube arrangement and whiteboard wiping, but fails on some object-generalization and inverse-action tasks. Astra is substantially stronger in spatial understanding, instruction following, and compositional generalization, while still failing some inverse tasks such as removing a bowl from a microwave or unfolding a rag. These failures reinforce the paper’s argument that zero-shot robot use is not equivalent to uniform physical competence.

Capabilities associated with effective robot use

The paper compares model generations, model families, and model sizes under a common harness. The results reject simple explanations based only on recency or parameter count. GPT-5.6 Sol already reaches 19.17% success and a 24.04 score, while GPT-6.1 Sol reaches 37.50% and 45.63. In the Qwen family, earlier Qwen3.5–3.7 models record zero success, whereas Qwen3.8-27B reaches 8.06% and Qwen3.8-Flash-Next reaches 25.00%. The 27B Qwen3.8 model succeeds where earlier models exceeding 150B parameters fail.

The resulting claim is deliberately stronger than “larger models perform better”: training distribution and acquired capabilities appear more predictive than parameter count alone. This conclusion remains correlational because the generations differ in architecture, optimization, data, prompting, and inference configuration. Nevertheless, the cross-family pattern is inconsistent with model size or release recency as sufficient explanations.

Trajectory analysis identifies four recurring failure classes:

1. General visual understanding: models misinterpret object identity, orientation, digits, or action outcomes. Earlier Qwen models exhibit such errors in 12.2–23.8% of trials, compared with 3.2% for Qwen3.8-27B and none in the more successful Qwen3.8-Flash and GPT-5.6 Sol cohorts.
2. Spatial grounding: models select inconsistent corresponding points across views, producing large reprojection errors and rejected triangulation calls. Even GPT-6.1 Sol has a 31.5% spatial-grounding failure rate under the paper’s measurement, indicating that fine-grained 3D correspondence remains a major bottleneck.
3. Geometric coding: models make errors in coordinate transforms, quaternion conventions, and orientation construction. Earlier Qwen cohorts show errors in 47.3–81.8% of executed geometric calculations, compared with 8.2% for Qwen3.8-Flash, 1.5% for DeepSeek, 0.7% for GPT-5.6 Sol, and none for GPT-6.1 Sol.
4. Embodied reasoning: models propose infeasible grasp orientations or physically inappropriate interaction strategies. GLM exhibits a 26.2% error rate in the narrow grasp-pose analysis, while GPT-6 Astra still encounters difficulties with bimanual coordination and other interaction structures.

The paper’s interpretation is that robot use emerges from the joint availability of these capabilities. A model may understand an object visually but fail to ground it in 3D; it may compute a pose correctly but select an infeasible grasp; or it may execute a valid grasp without recognizing that placement failed. This decomposition is more informative than treating robot-use performance as a monolithic “embodiment” score.

Figure 7: Lower error rates in visual understanding, spatial grounding, geometric coding, and embodied reasoning are associated with higher RoboDojo scores.

Correlations with general benchmarks support this interpretation. RoboDojo performance is positively associated with MMMU-Pro for visual understanding (Yue et al., 2024), ERQA for embodied reasoning (Team et al., 25 Mar 2025), ScreenSpot-Pro for spatial grounding (Li et al., 4 Apr 2025), and LiveCodeBench for code generation (Jain et al., 2024). The spatial-grounding correlation is weaker, plausibly because ScreenSpot-Pro evaluates 2D GUI grounding rather than 3D physical correspondence. These results are suggestive rather than causal: benchmark performance may reflect shared latent capabilities without guaranteeing transfer to robot interaction.

Figure 8: RoboDojo scores rise with several general model capabilities, with particularly strong relationships for visual understanding and embodied reasoning.

Geometric information, skills, and agent-written code

The depth ablation distinguishes three observation regimes: RGB plus depth and camera parameters; RGB plus camera parameters; and RGB alone. Removing depth while retaining camera calibration changes Astra’s score from 55.83 to 55.04, preserving 98.6% of the depth-enabled score. Qwen’s score decreases from 35.89 to 31.39, preserving 87.5%. Removing camera parameters as well reduces Qwen to 9.13 and Astra to 31.29.

The implication is specific: depth is not necessary for strong robot use when camera calibration is available, but calibration is substantially more valuable than RGB alone. Models can use image correspondences, camera intrinsics, and extrinsics to triangulate 3D locations or test geometric hypotheses. This does not mean that geometric information is unnecessary; it means that explicit depth can be replaced by other structured geometric information at a performance cost that depends strongly on model capability.

The paper provides concrete examples of Qwen writing and executing its own geometric code. In insert_tubes, it selects corresponding pixels across camera views, intersects viewing rays, and verifies reprojection errors as low as 0.46 and 1.94 pixels. In cover_blocks, it varies candidate table heights, computes ray-plane intersections, estimates cup dimensions, and tests scene hypotheses before acting. These examples demonstrate that the agent is not merely invoking a fixed perception primitive; it is using programmatic computation as part of spatial reasoning.

Figure 9: Agent-written triangulation converts multi-view pixel correspondences into 3D tube positions and validates them through reprojection error.

Figure 10: The agent tests candidate geometric hypotheses, including table height and cup dimensions, before committing to manipulation.

The skill-library ablation produces a smaller effect. With skills enabled, Qwen reaches a score of 31.39 and a 25.00% success rate. Without task-specific skills and skill writing, it reaches 30.31 and 24.58%, retaining 96.6% of the score and losing only 0.42 percentage points in success rate. This result supports the claim that the main performance does not depend on encoded task-specific solutions. It does not, however, show that skills are generally unimportant: the experiment retains standard tools, task instructions, and environment descriptions, and tests only one model-harness configuration.

Figure 11: Removing depth while retaining camera parameters causes only a modest performance decrease, whereas removing camera parameters produces a much larger decline.

Figure 12: Disabling the task-specific skill library produces only a small aggregate performance loss for Qwen3.8-Flash-Next.

Cost and throughput trade-offs

The resource analysis exposes a trade-off between absolute performance, cost, token consumption, and wall-clock time. Qwen achieves a RoboDojo score of 31.39 at an estimated $1.09 per episode, whereas Astra reaches 51.90 at$21.29 per episode, nearly 20 times the cost. DeepSeek lies between these configurations in performance but is less favorable in the reported resource comparison.

The token-budget curves complicate a simple “cheaper is faster” interpretation. To reach a score of 25, Astra requires 7.4 million tokens per episode, compared with 23.5 million for Qwen and 49.9 million for DeepSeek. In wall-clock time, Astra reaches that score after 42.6 minutes, while Qwen and DeepSeek require 95.5 and 99.2 minutes, respectively. Qwen is therefore much cheaper per episode but not more efficient in tokens or elapsed time in this experimental setup.

Figure 13: Qwen offers the lowest reported cost per episode, while Astra provides the highest score at substantially greater expense.

Figure 14: Models exhibit distinct progress curves as a function of token usage and elapsed time.

These measurements depend on provider pricing, reasoning settings, caching behavior, retries, and API conditions. They should be understood as deployment-cost estimates for the specified configurations rather than intrinsic model costs. Their principal significance is methodological: access to open-weight models can reduce the financial barrier to collecting large robot-interaction datasets, even when the open model is not the most token- or time-efficient policy.

Limitations and open questions

The paper’s strongest claims are subject to several limitations. First, the headline open-versus-closed comparison is not a matched model comparison: Qwen3.8-Flash-Next with uni-robot-agent is compared against a published Astra–RoboProbe result, while Astra evaluated with the proposed harness scores much higher. The harness contribution is supported by direct Qwen ablations, but the relative contributions of model, prompting, reasoning budget, tool design, and evaluation implementation are not fully disentangled.

Second, the real-world evaluation gives uni-robot-agent access to depth while comparison systems use RGB, and the benchmark uses rubric-based partial credit rather than a common full-task success measure. The physical results are consequently evidence of practical zero-shot capability, not a controlled demonstration of superiority over $\pi_{0.5}$ or Dream-1.

Third, the capability analysis is observational. Correlations between RoboDojo and general benchmarks do not establish that improving MMMU-Pro, LiveCodeBench, or ERQA will causally improve robot use. The embodied-reasoning analysis is also narrow, relying substantially on grasp-pose assessment and a limited set of trajectories. Bimanual coordination, contact-rich manipulation, force control, and recovery under severe visual ambiguity remain incompletely characterized.

Fourth, several comparisons use different seeds, layouts, reasoning efforts, model access modes, or evaluation budgets. The main RoboDojo result uses six episodes per task, whereas supplementary model comparisons use one seed and one layout. Pipelining results are based on one episode per task and configuration. These choices are reasonable for an engineering study but limit statistical certainty about smaller performance differences.

Finally, the harness is evaluated primarily on tabletop manipulation and a specific simulator/robot stack. Whether its abstractions remain adequate for mobile manipulation, deformable objects, force-sensitive tasks, dynamic environments, or heterogeneous multi-robot systems remains open. The paper specifically leaves unresolved how to train open-weight models on the structured interaction records that the harness can collect, and whether joint training of visual, geometric, coding, and embodied capabilities will improve long-horizon reliability without increasing inference cost.

Conclusion

The paper establishes uni-robot-agent as a unified RUA harness whose principal contribution is architectural rather than algorithmically specialized. Shared tools, explicit adapters, snapshot-consistent observations, diagnostic feedback, monitored execution, context management, and resource-aware pipelining provide a common substrate for open and closed models across simulation and physical robots.

Its empirical results show that Qwen3.8-Flash-Next can achieve a RoboDojo score of 31.39 without privileged simulator inputs, exceeding the published Astra–RoboProbe score of 28.97, while obtaining a real-world score of 49.50 across 19 tasks. The broader analysis indicates that effective robot use depends on a joint profile of visual understanding, 3D grounding, geometric computation, and embodied reasoning. The remaining question is not merely whether open models can issue robot commands, but how the structured interaction data enabled by this harness can be used to train those capabilities while preserving reliability, affordability, and execution efficiency (2611.01560).