Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation

Simulation has long been where robots learn to walk, roll and hop; at the Conference on Robot Learning (CoRL) in Austin, Texas, that proving ground is being repurposed for work needing a finer touch — aligning gears, connecting cables and assembling small components. NVIDIA is using the event to introduce open robotics technologies spanning world action models, simulation tools, benchmarks and agentic workflows, with the shift from locomotion to manipulation as the throughline.

The program also spotlights what organizers call agentic robotics, a paradigm that brings frontier agents in to generate interpretable code for physical robotic systems. Two NVIDIA papers are among six finalists chosen from thousands of submissions for CoRL's Best Paper Awards: GRAIL, which describes a digital generation pipeline producing 3D assets and simulator-ready scenes for robot learning, and RoboTTT: Context Scaling for Robot Policies, a model and training recipe that scales visuomotor context three orders of magnitude beyond state-of-the-art policies while keeping inference latency from growing.

The stated goal of these systems is to let researchers train robots on complex tasks, evaluate them across changing conditions and feed failures back into learning — the precision, adaptability and autonomy that separate lab demonstrations from industrial deployment. NVIDIA frames world action models as the capability that makes that possible, since the next level of autonomy requires robots to perceive surroundings, anticipate how their actions will alter the environment and carry out complex multistep tasks precisely.

Headlining the releases is GR00T 2, an open frontier world action model built on NVIDIA Cosmos. Trained on diverse robot and human data — including 100,000 hours of NVIDIA open source human egocentric footage — the model is aimed at dexterous manipulation skills. On the reported benchmark, it completed 40 of 50 evaluated tasks at least once without needing task-specific demonstrations, a signal of out-of-the-box generalization rather than per-task tuning.

The work does not arrive in isolation: NVIDIA names partners including Ambi Robotics, Bosch Research, Dexmate, Dyna Robotics, FieldAI, Flexion, Noitom Robotics, Siemens, Skild AI, RLWRLD, UC Berkeley, Walden Robotics and the University of Washington as collaborators on bringing these technologies to the wider robotics community.

Open questions center on durability and transfer. Completing 40 of 50 tasks once is a threshold, not a reliability rate, and scaling visuomotor context in training has to survive contact-rich work where tolerances are measured in millimeters. Whether agentic code generation holds up outside demos, and how quickly GR00T 2's open release translates into deployed manipulation systems, are what to watch after Austin.