R1: Sim‑Only Manipulation
- sudo Robotics demonstrated an R1 manipulation model trained purely in simulation that generalises to novel real objects. - The model reportedly handles fragile glass and deformable fabrics without any real‑world training data. - Simulation‑only training reduces real‑world data needs and could accelerate manipulation research and deployment pipelines. (x.com)
Robots usually learn to pick things up from real-world demos because contact is messy; sudo Robotics said its new R1 learned that skill in simulation alone. (sudo.ai) The company said R1 was trained with “zero real-world data” and then tested on unseen physical objects, including transparent glass, soft fabric, reflective metal, and irregular items. It reported a 60-minute uncut evaluation with about 98% first-attempt success and nearly 100% success within two tries. (sudo.ai) R1 is built around object picking, the first move behind many warehouse, factory, and household tasks. sudo Robotics said the system runs fully closed-loop at 15 to 25 hertz, meaning it updates its motion several times a second from fresh camera input instead of following a fixed plan. (sudo.ai) That distinction matters in manipulation because grasping is where robots still fail on new objects, glare, clutter, and small bumps. sudo Robotics said R1 kept working under changing lighting, dynamic backgrounds, random physical interference, and obstacle-constrained placements. (sudo.ai) Simulation has been a robotics goal for years because real data collection is slow, expensive, and often hard to scale across edge cases. Stanford’s BEHAVIOR benchmark, built on NVIDIA’s Omniverse-based OmniGibson simulator, reflects that push with 1,000 household tasks and support for cloth, fluids, and other complex interactions. (behavior.stanford.edu) The hard part is the “sim-to-real” gap: a policy that works in a virtual scene can fail when a real camera sees glare, a real gripper brushes fabric, or a real object flexes. sudo Robotics said closing that gap required physics fidelity, contact modeling, domain randomization, and sensor simulation “simultaneously.” (sudo.ai) The company has not, at least on its public page, released a peer-reviewed paper, full benchmark protocol, or outside replication of the R1 results. Its website presents the system as a self-developed hardware-and-software stack and frames the demo as a step toward “production-grade performance” rather than proof that general robot manipulation is solved. (sudo.ai) That caution lines up with the wider robotics market. Boston Consulting Group wrote on April 14, 2026, that dexterous manipulation is improving, but leaders still need to separate “deployable capabilities” from polished demonstrations as capital pours into physical artificial intelligence. (bcg.com) For now, the claim is narrower and concrete: a robot trained in virtual worlds picked up unfamiliar real objects, including fragile and deformable ones, without real-world demos. If other labs can reproduce that result, one of robotics’ slowest bottlenecks could start to move faster. (sudo.ai)
