Robots rehearse delicate assembly in simulation: NVIDIA's Isaac Lab 3.0 with Newton reaches up to 2.5x training throughput

The simulation focus in robot learning is shifting from locomotion such as walking, rolling and hopping toward manipulation tasks that need a finer touch, including aligning gears, connecting cables and assembling small components, and NVIDIA's open tools released at CoRL target that shift. Simulation had long served as the proving ground for robots learning locomotion, while fine manipulation places different demands on physics accuracy and data scale. NVIDIA introduced world action models, simulation tools, benchmarks and agentic workflows together, working with partners including Ambi Robotics, Bosch Research, Siemens and UC Berkeley to bring them to the broader robotics community.

Isaac Lab 3.0 lets researchers use agents to set up training environments quickly, switch between the PhysX and Newton physics solvers, and apply a new reinforcement post-training library to foundation models such as GR00T. Tighter integration with Newton lets Isaac Lab model deformable objects such as cables and cloth, along with complex articulated systems such as robotic hands with many interacting joints. On a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX, and the combination can train thousands of simulated robots in parallel.

AlphaNIST, built in Isaac Lab, learns all 20 tasks on the NIST Assembly Task Board with a single policy, manipulating millimeter-scale parts using only a yes-or-no success signal. Separate per-task policies reached 81% training success on the same task board. AlphaNIST achieved more than 98% training success using roughly 39x less GPU time; its Success-Guided Sampling focuses practice on tasks a policy succeeds at about half the time, letting training scale beyond a million parallel environments.

The demonstrated scope centers on training and evaluation in simulation: the throughput comparison for Isaac Lab 3.0 with Newton comes from a vision-based dexterous manipulation benchmark, the AlphaNIST result covers the 20 tasks on the NIST Assembly Task Board, and the GR00T 2 success rates come from 50 evaluated tasks. The intended readers are robot-learning researchers and engineering teams, and uses of GR00T 2 by Nexuni, Foxconn and Siemens are at testing, application or evaluation stage.

The 71% and 95% average success rates for GR00T 2 correspond to one and five demonstrations per task respectively, so they are results under different data conditions; Isaac Lab-Arena 0.4 arrives as a developer preview; and some DexBench evaluators, including Genesis AI and mimic robotics, are in early access.