sudoAI's goal is to build a foundation for general-purpose robots so that robot skills can be learned, transferred, and deployed in the physical world. Its launched sudo R1 is an integrated robot system with both hardware and software fully developed in-house, driven by a manipulation-centric foundation model focused on object picking. Picking was chosen as the entry point because almost every useful manipulation task begins with picking, and if picking cannot work reliably on long-tail objects in the real world, many downstream tasks are difficult to achieve.
In terms of main capabilities, sudo R1 has the following features. First, zero-shot generalization: it can successfully pick a diverse range of objects never seen in training, covering rigid and deformable, opaque and transparent, matte and reflective types, including transparent glass, soft fabrics, reflective metal, and irregularly shaped items. A single model can handle all of these cases without fine-tuning and without adaptation for individual objects. Second, robustness: under controlled lighting changes and with a TV screen used to simulate various dynamic backgrounds, the picking success rate remains nearly consistent, without environment-specific calibration or fine-tuning. This robustness comes from large-scale visual randomization during simulation training, causing the policy to rely on geometry and physical cues related to grasping rather than overfitting to a specific visual environment. Third, true closed-loop agility: sudo R1 uses a fully closed-loop policy, where each control step is conditioned on the robot's latest observation and runs at an adaptive frequency of 15–25 Hz. Compared with models that use action chunking and execute a fixed action sequence between two observations, this design enables continuous feedback during execution, thereby supporting tracking moving targets, recovering after being disturbed mid-grasp, and adjusting motion trajectories when the scene changes. This is also why the system can run at speeds close to production relevance. Fourth, spatial intelligence: when surrounding objects or structures restrict feasible paths, sudo R1 can adjust trajectories, bypass obstacles, avoid collisions, and use available space, and this is an integrated capability of the learned policy rather than an added independent collision avoidance module.
In terms of training approach, sudo R1 is trained entirely on simulation data, without using real-world demonstrations, teleoperation, or human annotation. Its capability improvement comes from generating more data rather than simply scaling up manpower, forming a different scaling curve. The team believes simulation is key because achieving requirements simultaneously across the four dimensions of generalization, dexterity, robustness, and high-frequency control is limited by the data bottleneck: relying only on real-world collection is too slow, too expensive, and too narrow in coverage, making it difficult to systematically construct adversarial conditions and generate dense obstacle scenarios; simulation can scale across all dimensions simultaneously. Achieving this requires bridging gaps in every part of sim-to-real, including physics fidelity, contact modeling, domain randomization,and sensor simulation, and these capabilities have been built through years of dedicated engineering.
In terms of applicable scenarios, the official text mentions that physical AI has economic potential in fields such as manufacturing, logistics, agriculture, and elderly care, but current manipulation capabilities are still fragile. sudo R1 targets picking as an entry-level action and plans to expand to more skills.
In terms of limitations, the official statement explicitly says that true production-grade performance still requires time: achieving any one of generalization, agility, robustness, or spatial intelligence alone is already difficult, while achieving them simultaneously in a single policy is a fundamentally different challenge. This is the goal sudo R1 pursues, but it has not yet been achieved. Picking is only the beginning, and other skills are still being expanded.
Other information such as specific pricing, charging model, platform support, installation or usage commands, and technical documentation is not provided in the official text.
