NVIDIA, Partners Show Robots Rehearsing Delicate Tasks in Simulation at CoRL

At the Conference on Robot Learning in Austin, Texas, NVIDIA and industry partners showcased advances that push robotics from locomotion toward precise manipulation and deployable industrial systems. The companies introduced open world action models, simulation frameworks, evaluation benchmarks and agentic workflows designed to help robots learn delicate tasks such as gear alignment, cable connection and small-component assembly. Two NVIDIA papers were among six CoRL Best Paper Award finalists from thousands of submissions. A central announcement is GR00T 2, an open frontier world action model built on NVIDIA Cosmos and trained on diverse robot and human data, including 100,000 hours of open-source egocentric human footage. The model is intended to give robots stronger perception, prediction and multistep task execution, particularly for dexterous manipulation. It ranks among the top models on RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks. Nexuni is testing it for robotic laundry handling, Foxconn is applying it to server-manufacturing automation, Siemens is evaluating it for industrial use cases, and MicroAGI is using it for Ethernet cable swapping with a YAM robotic arm. NVIDIA also emphasized simulation as the training and scaling layer. Isaac Lab 3.0, an open robot-learning framework, lets researchers use agents to configure training environments, switch between NVIDIA PhysX and Newton physics solvers, select renderers for perception-based training and apply reinforcement-learning post-training to foundation models such as GR00T. Deeper integration with Newton improves simulation of deformable objects such as cables and cloth, as well as complex articulated systems such as robotic hands. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton delivered up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Partners are extending those tools. Skild AI introduced STEEL, a rigid-body physics solver built on Newton to accelerate training with Isaac Lab. Flexion is using Newton and Isaac Lab to train its Reflect model, which combines reasoning and whole-body manipulation. FieldAI is training and validating policies for unstructured real-world environments, while Dyna Robotics is using parallel simulation to train whole-body controllers for its semi-humanoid Taku robot. Walden Robotics is using photorealistic simulation for general-purpose robots in manufacturing and logistics, and Dexmate is exploring coding agents that refine manipulation strategies before deploying them on its Vega robot. One result illustrating the impact of efficient simulation practice is AlphaNIST, built in Isaac Lab. Using Success-Guided Sampling from the University of Washington and NVIDIA, a single policy learned all 20 tasks on the NIST Assembly Task Board with more than 98 percent training success, compared with 81 percent for separate per-task policies, while using roughly 39 times less GPU time. Evaluation advances were also introduced. Isaac Lab-Arena 0.4, available in developer preview at CoRL, enables scaled policy evaluation across thousands of variations and helps identify conditions that cause failures. Two new open benchmarks, DexBench and VABAR, measure robot dexterity and adaptability in industrial settings. DexBench, developed by RLWRLD with NVIDIA, covers 15 tasks across retail, logistics, manufacturing and hospitality. VABAR, from UC Berkeley and NVIDIA, evaluates six industrial tasks using standardized variations, difficulty tiers and success metrics. For program generation, UC Berkeley’s Graph-as-Policy framework uses agents to connect individual robot skills into interpretable robot programs and refine them through simulated failures. Partners including Ambi Robotics, Bosch Research and Siemens are evaluating the approach for industrial automation. NVIDIA’s CoRL releases aim to make robot training, evaluation and program development more precise, scalable and ready for real-world deployment.