Jiaming Tang, Zhuoyang Zhang, Junxian Guo, Fangzhou Zhao, Yilong Zhao, Zekai Li, Xingyang Li, Yuyang Zhao, Song Bian, Zhijian Liu, Song Han
8 min
This paper explores whether general-purpose reasoning models, similar to computer-use agents, can be trained to control robots natively by observing scenes, reasoning in a chain of thought, and outputting end-effector poses. Specifically, the authors aim to bridge the performance gap between frontier proprietary models and open-weight models without relying on task-specific harnesses or human demonstrations.
To solve this, the authors built a Post-training Engine for Agentic Robotics (PEAR). They trained a 27B open-weight model using robot-use reinforcement learning (RURL) entirely in simulation, drawing on techniques from computer-use agents. The agent receives camera images and the robot's current state, produces a chain of thought, and outputs a target pose. Inverse kinematics then handles the conversion to joint commands. Training followed a multi-stage curriculum, starting with easier tasks in the LIBERO suite and progressing to more complex tasks in RoboDojo, allowing the agent to learn transferable spatial and manipulation skills.
The model shows remarkable zero-shot transfer capabilities. After training on simulation benchmarks (LIBERO and RoboDojo), the model achieved a 60% success rate on a real-world dual-arm robot that it had never encountered during training. Furthermore, it demonstrated significant robustness to task perturbations and novel instructions, outperforming established vision-language-action (VLA) models on benchmarks like LIBERO-Pro. Notably, this specialized training resulted in no measurable degradation of the model's performance on general reasoning, coding, and vision benchmarks, suggesting that robot-use training can be seamlessly integrated into a foundation model's post-training regimen.
This work provides a simple, scalable recipe for building robot-use agents that do not require fragile, task-specific software harnesses. By proving that robotic control is a skill that can be acquired by reasoning models through simulation-based reinforcement learning, the authors open a path toward creating truly general-purpose robotic agents that can handle diverse, real-world tasks simply by following natural language instructions.
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 $π_{0.5}$, a vision-language-action model fine-tuned on demonstrations, our model is substantially more robust to changes in the task itself. On LIBERO-Pro, $π_{0.5}$ 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.
Sam: In a real-world task, the instruction asks the robot to use a set square to push blocks into an aligned row. It must then release the tool and reset the arm. The model repeatedly observes, reasons, requests a move, and checks what happened.
Alex: How does it learn that sequence without demonstrations? A failed attempt doesn’t tell it which earlier decision went wrong.
Sam: They use reinforcement learning, meaning learning from rewards for attempted tasks. Here the reward is deliberately sparse: success or failure at the episode’s end. They compare groups of attempts at the same task and starting state, then reinforce responses associated with better outcomes.
Alex: But if the starting model almost never succeeds, that reward gives it very little to learn from. How do they get past that?
Sam: They start with simulated tasks where it already succeeds occasionally. Learning there transfers to harder tasks, which they add once success becomes frequent enough to provide a signal. Their training engine, PEAR, supports these extended interactions at scale.
Alex: Let’s separate learning the training environment from learning something more general. How large was the improvement on the main benchmark?
Sam: On LIBERO, a suite of simulated manipulation tasks, success rose from about two percent to seventy-four percent. That brought the model close to the strongest proprietary comparisons at their maximum reasoning settings. But performance on familiar tasks isn’t the most revealing comparison here.
Alex: Because a robot can score well by repeating the motion associated with a familiar scene. Did they test whether it actually follows a changed instruction?
Sam: They did, using LIBERO-Pro, which alters goals and layouts. Their comparison was a vision-language-action model, connecting images and instructions to robot actions, called pi point five. It had been trained with demonstrations of every original LIBERO task.
Alex: That gives the demonstration-trained model an advantage on familiar tasks. What happens when the scene stays familiar but the requested task changes?
Sam: The reasoning model succeeded about fifty-seven percent of the time. The demonstration-trained model succeeded less than one percent of the time. The authors report that it tended to perform the task originally associated with that scene instead.
Alex: That’s a stronger test of instruction following than merely changing the wording. Does it mean the reasoning model is more robust across the board?
Sam: No. The demonstration-trained model still led when instructions were reworded or objects restyled without changing the required task. It also led on the original benchmark. And those comparison results came from prior reports, not fresh evaluations by this team.
Alex: Now take us out of simulation. How much real-world evidence sits behind the opening result?
Sam: Each model received thirty trials on a real dual-arm robot absent from training. The tasks were object pickup and aligning blocks with the set square. The final model had trained on those task types in simulation, but used no real-robot training data.
Alex: So this tests familiar goals on unfamiliar physical hardware. It doesn’t establish that the robot can handle arbitrary new real-world jobs.
Sam: That’s the key boundary. My assessment is that this is persuasive evidence of transfer within those tasks, not broad deployment readiness. The small real-world evaluation matters more than the headline percentage alone.
Alex: The document also includes drawing and a chemistry task. Do those extend the case beyond the practiced goals?
Sam: They provide additional demonstrations, not a comparable success-rate evaluation. The robot draws a simple version of The Starry Night and selects phenolphthalein to test alkalinity. But the authors say these harder tasks required detailed descriptions of both the robot and the task.
Alex: What else would stop someone treating this as a ready-to-use robot controller?
Sam: Speed and prompt dependence. It reasons at length before every move, and lowering the reasoning effort sharply reduces success. It also cannot reliably plan chunks of actions, so complex movements require repeated decisions and stops. Controller quality still affects accuracy.
Alex: And does learning robot control displace the model’s other abilities? That matters if this is supposed to become general model training.
Sam: Across the tested reasoning, coding, and vision benchmarks, they found no measurable degradation. That supports the proposal, but doesn’t prove every other capability survives. Likewise, the effort comparison links longer reasoning with better performance; it doesn’t isolate why this method generalizes.
Alex: Who should read the full document, and where should they start?
Sam: Robot-learning researchers should start with the robot-use interface, then the robustness and instruction-following comparison. Foundation-model researchers should read the training curriculum and retained-capabilities evaluation. Both should read the sim-to-real section alongside the limitations, especially the narrow task coverage and dependence on detailed prompts.
Alex: And for someone who only needs the central lesson?
Sam: Simulated practice can teach a model skills that survive a change of robot. But transfer is not the same as general reliability.
Alex: Thanks for listening.