Robot-use agents (RUAs) have emerged as a promising approach to general-purpose robot control. However, strong performance has so far been achieved only by frontier proprietary models such as GPT-6 Astra and Claude Fable 5.2, whereas open-weight models remain substantially less capable. We show that an open-weight 27B reasoning model can be trained to use real robots natively through reinforcement learning, entirely in simulation and without demonstrations. Like a computer-use agent (CUA), the model observes the environment, reasons in a chain of thought, and generates the next end-effector pose, without any harness. We build a Post-training Engine for Agentic Robotics (PEAR) and use it to train the model on simulated robot tasks at scale with robot-use reinforcement learning (RURL), mirroring how frontier models acquire computer use. We find that the skills acquired in simulation generalize well beyond the training tasks: they transfer to an unseen robot, to the real world, and to novel instructions. After training on LIBERO and RoboDojo in simulation, the model succeeds in 60% of our trials on a real dual-arm robot, whereas the base model succeeds in none. On LIBERO, training raises the success rate from 1.6% to 74.4%, approaching GPT-6 Astra and GPT-6.1 Sol at maximum reasoning effort. Compared with π0.5, a vision-language-action model fine-tuned on demonstrations, our model is substantially more robust to changes in the task itself. On LIBERO-Pro, π0.5 succeeds in only 0.75% of episodes when assigned a new task in a familiar scene whereas our model succeeds in 57.4%. Finally, the model retains its general capabilities, with no measurable degradation on nine reasoning, coding and vision benchmarks. These results highlight robot-use reinforcement learning as a promising component of foundation-model post-training.
Figures & tables
Figure 2: One round of robot use. A computer-use agent (top) observes a screenshot, reasons about the task and outputs a click or a keystroke, which the operating system executes. Our model (bottom) operates a robot in a similar way: it observes the camera views and the arm’s state, reasons in a chain of thought and outputs the end-effector’s next pose, and inverse kinematics and the arm’s controller then move the arm. Each round is added to the context, which also holds the instruction and all earlier rounds. The same steps are used in simulation and on the real robot.
Figure 3: Training curve on LIBERO and comparison with frontier VLMs. (a) The success rate of training rollouts per RL step (thin line) and smoothed (thick line); LIBERO-Long is added at step 149, and the diamonds are the same run evaluated on LIBERO’s 2,000 episodes. (b) The base model, our model at xhigh, the effort used in training, and GPT-6.1 Sol ( OpenAI, 2026a ) and GPT-6 Astra ( OpenAI, 2026b ) at medium and max effort, on LIBERO’s 2,000 evaluation episodes; all models use the same protocol, and effort is each model’s own setting for how long it reasons.
Figure 4: Comparison with π0.5 on LIBERO, LIBERO-Plus and LIBERO-Pro. Both policies are trained only on LIBERO: π0.5 with supervised fine-tuning on demonstrations and our model with RURL. (a) Success rate on LIBERO and on the two perturbed benchmarks. (b, c) Two of LIBERO-Pro’s perturbations, an unseen task in a training scene and swapped object positions; the arrow gives the drop from LIBERO. The π0.5 results are from prior work ( Physical Intelligence, 2025 ; Yan et al., 2026 ; Zhou et al., 2025 ) .
Figure 5: Zero-shot success rate on unseen tasks, a new simulator and robot, and a new benchmark. (a) RL without LIBERO-Long is trained on the other three LIBERO suites; RL with LIBERO-Long, the same run after LIBERO-Long is added, is shown for reference. (b) RoboDojo, a different simulator with a dual-arm ARX, after RL on LIBERO only; error bars are 95% Wilson intervals. (c) RoboLab-120, a third benchmark, against four generalist robot policies from the RoboLab leaderboard ( NVIDIA, 2026c ) , all DROID ( Khazatsky et al., 2024 ) checkpoints.
Figure 6: Training curve on RoboDojo subset and success rate before and after RL. The final model continues from the LIBERO-only model with RURL on two RoboDojo tasks, entirely in simulation. (a) Success rate of training rollouts per RL step (thin line) and smoothed (thick line). (b) Success rate on the two tasks before and after this training; error bars are 95% Wilson intervals.
Figure 7: Real-robot evaluation on a dual-arm YAM. Our training does not include this robot and uses no real-robot data. (a) General pickup: three trials in which the final model picks up the object and the base model does not; the bar chart gives each model’s successful trials out of 20, two for each of ten objects. (b) Align blocks: three rounds of one trial in which the final model succeeds and the base model fails; the bar chart gives each model’s successful trials out of ten. (c) Reasoning tasks: the final model draws a simple line version of a painting and selects a reagent that shows whether a solution is alkaline.
Figure 8: General benchmarks before and after RL. The final model against the base model on the same questions: the change in score (dots) with its paired 95% interval (lines), and the scores before and after RL beside each name. The benchmarks are AIME 2026, GPQA Diamond, IFBench, LiveCodeBench v6 (LiveCode), RealWorldQA, ERQA, the MMMU validation set and MMMU-Pro in its standard (S) and vision (V) settings; the mean change over the nine is −0.15 points. Table 6 gives the full numbers and the evaluation settings.
Figure 9: Reasoning effort and inference speed. (a, b) Thinking tokens per round and success rate on LIBERO of the LIBERO-only model at low and xhigh effort, with BF16 weights; the low-effort values are the mean of two evaluation runs. (c) Decoding speed on one device, with NVFP4 weights quantized with NVIDIA ModelOpt ( NVIDIA, 2024 ) and Inco’s DFlash2 drafter throughout; our megakernel on the GPUs and Inco’s Splash engine on the M5 Max.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Value
Base model
Qwen3.8-27B ( Qwen Team, 2026 )
RL algorithm
GRPO ( Shao et al., 2024 ) with DAPO’s group filter ( Yu et al., 2025 ) and score centering ( Marek & Ryabinin, 2026 ) ; a REINFORCE ( Williams, 1992 ) policy loss in bypass mode, without PPO clipping ( Schulman et al., 2017 )
Batch
128 groups of 8 rollouts per step, counted after the group filter
Optimization
AdamW ( Loshchilov & Hutter, 2019 ) with a learning rate of 10−6 and one update per step; the loss is averaged over the generated tokens of a batch (token-mean), without a KL penalty or an entropy bonus
Sampling
Temperature 1.0 over the full distribution
Curriculum
LIBERO: Goal, Spatial and Object for steps 1–148; all four suites from step 149 to step 225 (the LIBERO-only model). RoboDojo: general_pickup and align_blocks for 115 steps from the LIBERO-only model (steps 226–340)
Appendix
Table 1: Training setup.
Model
Effort
Spatial
Object
Goal
Long
Average
Vision-language models
Qwen3.8-27B
xhigh
1.2
1.8
2.8
0.4
1.6
GPT-6.1 Sol ( OpenAI, 2026a )
max
78.8
83.4
76.8
65.8
76.2
GPT-6 Astra ( OpenAI, 2026b )
max
82.6
82.2
80.2
77.4
80.6
GPT-6.1 Sol
medium
22.8
69.2
58.2
42.2
48.1
GPT-6 Astra
medium
26.0
67.0
59.4
44.6
49.3
Appendix
Table 2: LIBERO success rate (%) by suite. The low-effort row is the mean of two evaluation runs. The π0.5 row is from the openpi repository ( Physical Intelligence, 2025 ) . Bold: best VLM in each column, among the max and xhigh rows and among the medium rows.
Figure 10: LIBERO by suite, and LIBERO-Pro and LIBERO-Plus by perturbation. (a) Success rate (%) on LIBERO by suite, against the frontier VLMs at max effort ( Table 2 ). (b) Success rate (%) on LIBERO-Pro by perturbation type: an unseen task in a training scene, swapped object positions, a reworded instruction and restyled objects ( Table 3 ). (c) The LIBERO-only model on LIBERO-Plus by perturbation dimension (10,030 episodes); the dashed line marks its success rate on LIBERO.
Model
Task
Position
Semantic
Object
Mean
π0.5 , supervised fine-tuned on LIBERO
0.75
20.8
95.8
96.0
53.3
Ours (LIBERO only), xhigh effort
57.4
57.5
73.3
60.2
62.1
Appendix
Table 3: LIBERO-Pro success rate (%) by perturbation type , both policies trained on LIBERO; 2,000 episodes per type for our model. The π0.5 row averages the per-suite results of Zhou et al. (2025, Tables 2–5 of the first arXiv version) over the four suites. Bold: best in each column. Task: the same scene with a different task. Position: the objects’ positions swapped. Semantic: the instruction reworded. Object: the objects restyled. Mean: over the four types.
Suite
LIBERO
LIBERO-Plus
LIBERO-Pro
Spatial
85.8
75.7
78.9
Object
97.0
82.7
86.4
Goal
61.4
46.8
48.7
Long
53.2
44.5
43.8
Appendix
Table 4: The LIBERO-only model by suite, on LIBERO and on the two perturbed sets (success rate, %); perturbations reduce every suite’s success rate by 7–15 points. LIBERO-Pro’s numbers include its unperturbed tasks (64.5 on average; 62.1 over the four perturbations alone). LIBERO-Pro results by perturbation type are in Table 3 , and LIBERO-Plus results by perturbation dimension in Figure 10 c.
Model
All
Simple
Moderate
Complex
Score
Our base model
Qwen3.8-27B
5.4
5.2
7.7
1.2
12.1
Generalist robot policies
Cosmos3-Nano-Policy ( NVIDIA, 2026a )
36.8
40.6
35.4
25.3
51.9
π0.5 ( Black et al., 2025a )
28.0
29.7
31.5
13.5
43.4
DreamZero ( Ye et al., 2026 )
25.7
26.1
30.0
14.1
39.8
Appendix
Table 5: RoboLab-120 by task difficulty. Success rate (%) on all 120 tasks and by difficulty (64 simple, 39 moderate and 17 complex tasks, ten episodes each), and Score, RoboLab’s graded measure of progress through a task’s subtasks, where a success scores 100. The final model is trained with RL on LIBERO and RoboDojo, and its values are the mean of two evaluation runs (2,400 episodes). Other policies: the RoboLab leaderboard ( NVIDIA, 2026c , data of 29 October 2026) . Bold: best in each column.
Figure 11: Average success rate on the full RoboDojo benchmark. All four models are scored with RoboDojo’s official rule, the mean success rate of its five capability dimensions. The base model and the final model are evaluated with 10 episodes per task setting. GPT-5.5 and DeepSeek-Flash are entries of the leaderboard ( RoboDojo Team, 2026 ) from the RoboDojo team’s evaluation of language models as policies ( Zhang et al., 2026b ) , with 50 and 10 episodes per task.
Benchmark
Category
Questions
Base
Ours
Change
95% CI
AIME 2026
math
30 × 16
98.3
98.5
+0.2
[ −1.7 , +2.9 ]
GPQA Diamond
science
198 × 8
89.6
89.7
+0.1
[ −1.0 , +1.2 ]
IFBench
instruction following
300, prompt strict
73.0
72.7
−0.3
[ −4.0 , +3.3 ]
LiveCodeBench v6
coding
131 × 4
88.0
88.4
+0.4
[ −2.1 , +2.9 ]
RealWorldQA
visual QA
765
84.1
83.7
−0.4
[ −2.4 , +1.6 ]
ERQA
embodied reasoning
400
68.5
66.5
−2.0
[ −6.2 , +2.2 ]
Appendix
Table 6: General benchmarks. The final model and the base model on the same questions, with paired 95% intervals and the number of questions (and samples per question). The intervals are from a paired bootstrap over questions with 2,000 resamples. evalscope ( ModelScope Team, 2024 ) 1.13.0, vLLM ( Kwon et al., 2023 ) in BF16, thinking-mode sampling (temperature 1.0, top- p 0.95, top- k 20). Changes and means are computed before rounding. Nine-benchmark means: 81.64 for the base model and 81.49 for the final model. For reference, the LIBERO-only model averages 81.31.
Qwen3.8-27B (base)
Ours (LIBERO only)
Ours
Scene or object
Trial
Result
Rounds
Context
Result
Rounds
Context
Result
Rounds
Context
Align blocks
Scene 1
1
F: unaligned
12
40.0k
F: unsafe
20
78.2k
S
10
54.8k
2
F: rounds
25
77.7k
S
24
57.8k
F: unaligned
9
34.5k
Scene 2
1
F: rounds
25
103.4k
F: 32k tokens
10
99.4k
S
14
47.1k
2
F: rounds
25
60.6k
F: rounds
25
106.0k
S
8
26.1k
Appendix
Table 7: Every real-robot trial , all models at xhigh effort on a dual-arm YAM. Align blocks: two trials in each of five scenes, which differ in the initial layout of the blocks and the set square, all with the instruction “Use the set square to push the three blocks into a straight, aligned row, then release the set square and reset the robot arm.” General pickup: two trials per object, “Pick up the ⟨ object ⟩ by 10 cm.” S : success within the budget of 25 rounds; F: failure, followed by its cause. Rounds : ran out of rounds; 32k tokens : reasoned for more than 32k tokens in a single round; toppled : knocked the object over; unaligned : finished with the blocks not aligned; not lifted : did not fully lift the object; wrong object : picked up the wrong object; rolled : the ball rolled away; unsafe : stopped automatically in an unsafe pose. Rounds: the number of rounds, at most 25. Context: tokens used by the last of these rounds, out of 260,096.
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.
Frontier models such as GPT-6 Astra now directly control robots and can generalize better than vision-language-action models (VLAs). We argue that this ability could arise from two distinct, composable skills: semantic reasoning, to understand what a task asks for in a scene, and motion control, to move the robot to do it. We show that small open VLMs are already capable of the first and mainly lack the second. We introduce Zero-real Agent Policy (ZAP), a recipe that teaches this missing motion skill, entirely in simulation on simple primitive tasks and from the model's own successes, with no frontier teacher or distillation and no real-world robot data. ZAP reaches 96.7% success on real-world tasks, matching GPT-6 Astra in our harness while running 3x faster and 17x cheaper, and 74.0% on a bimanual adaptation of RoboLab-120, on par with GPT-6 Astra's 72.2%. ZAP also generalizes to new objects, goals, and strategies, and in simulation to a robot embodiment on which it never trained.
Benjamin Pikus, Daniel Saunders, Mircea Mironenco +11
GPT-6 Astra can drive a robot arm directly through a fixed harness, and scores of such robot-use agents on RoboDojo have risen quickly since. Reported systems differ in harness and test subset and often change several designs at once, so it is hard to tell which choices produce the gains. We take these designs apart and test them in one framework, along the harness around the model, the action interface, the reuse of experience and the scene geometry given to the model. A lean harness that keeps the full episode history and stops each motion early goes a long way. Reusing experience changes how the model acts but is not a free lunch, and scene geometry helps most, even when depth is predicted from RGB alone. JustRUA starts from the official L3 harness and keeps only what matters most. Even with no learned policy, self-improvement loop or task-specific tuning, this simple harness raises the score on the RoboDojo simulation benchmark substantially, from 29.7 to 50.3.
Zefan Wang, Yufan Dang, Xuanwei Shi +9
Tsinghua University · ModelBest · Huazhong University of Science and Technology +2