NVIDIA Unveils Open Robotics Technologies and GR00T 2 at CoRL

NVIDIA and various industry partners have unveiled a suite of open robotics advancements at the Conference on Robot Learning (CoRL) in Austin, Texas [1]. The announcements highlight a growing focus on robot manipulation tasks—such as gear alignment and cable connection—powered by simulation tools, benchmarks, and agentic workflows [1].

Advancements in Robot Autonomy and World Action Models

NVIDIA introduced GR00T 2, an open frontier world action model built on NVIDIA Cosmos and trained on diverse human and robot datasets, including 100,000 hours of NVIDIA open-source human egocentric data [1]. The model is designed to facilitate dexterous manipulation skills and features out-of-the-box performance, skill transfer between different robot bodies without requiring robot demonstrations, and generalisation with limited data [1]. According to NVIDIA, GR00T 2 completed 40 out of 50 evaluated tasks at least once without task-specific demonstrations and achieved a 71% average success rate across 50 tasks with a single demonstration per task, which rose to 95% with five demonstrations [1].

Several companies are already testing or applying the technology. Nexuni is testing it for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it for industrial use cases [1]. Additionally, MicroAGI uses GR00T 2 with reinforcement learning to perform Ethernet cable swapping with a YAM robotic arm [1]. Two NVIDIA papers also secured finalist spots for Best Paper Awards at CoRL: GRAIL, focusing on a digital generation pipeline for 3D assets, and RoboTTT, which introduces context scaling for robot policies [1].

Simulation, Isaac Lab 3.0, and Physics Engines

To bridge the gap between simulation and deployment, NVIDIA released Isaac Lab 3.0, an open-source robot-learning framework [1]. Isaac Lab 3.0 enables agents to rapidly configure training environments, switch between NVIDIA PhysX and Newton physics solvers, and apply a reinforcement post-training library to foundation models like GR00T [1]. Tighter integration with the open-source physics engine Newton allows the platform to model deformable objects like cables and cloth alongside complex articulated systems [1].

In benchmark testing, Isaac Lab 3.0 combined with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark [1]. Developers are actively expanding these capabilities. Skild AI introduced STEEL, a rigid-body physics solver built on Newton, while Flexion, FieldAI, Dyna Robotics, Walden Robotics, and Dexmate are utilising Isaac Lab, Isaac Sim, and Newton to train, validate, and simulate robot policies for warehousing, manufacturing, and logistics [1].

Evaluation Frameworks and Benchmarks

Evaluating robot reliability across changing conditions is being addressed through new tools such as NVIDIA Isaac Lab-Arena, a framework for scaled policy evaluation [1]. A developer preview of version 0.4 was released at CoRL [1]. Furthermore, NVIDIA and RLWRLD introduced DexBench, an open benchmark featuring 15 tasks across four scenes spanning retail, logistics, manufacturing, and hospitality to measure robot dexterity and adaptability [1].

The Conference on Robot Learning (CoRL) serves as a major venue for researchers and technology companies to showcase advancements in robotics, machine learning, and simulation [1]. As physical AI and robotics advance, simulation has become critical for training robots safely and efficiently at scale before physical deployment [1].

These developments provide researchers and industrial firms with open-source tools, models, and workflows to accelerate the training and deployment of autonomous robots for manufacturing, logistics, and retail. The integration of advanced simulation physics and generalisable action models aims to reduce the time and GPU resources required for complex robot training.

The availability of open-source world action models and advanced simulation frameworks like Isaac Lab 3.0 could significantly lower the barrier to entry for developing dexterous industrial robots. Enhanced simulation-to-real transfer efficiency may accelerate commercial deployments in warehousing, manufacturing, and logistics.

While simulation throughput and out-of-the-box success rates in virtual environments are high, the reliability of transferring these complex skills directly to unpredictable real-world industrial settings remains an ongoing engineering challenge.