NVIDIA unveils GR00T 2 and Isaac Lab 3.0 at CoRL 2026 to advance robot training
NVIDIA 在 CoRL 展示 GR00T 2 和 Isaac Lab 3.0，以推动先进机器人训练
At the 2026 Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA introduced GR00T 2, a world action model built on Cosmos utilizing 100,000 hours of open-source egocentric data, and Isaac Lab 3.0, an updated simulation framework featuring enhanced physics integration with Newton and improved training throughput. The company highlighted that GR00T 2 achieved 50% success on 40 tasks without specific demonstrations and 95% success with five demonstrations per task across 50 tasks. Partnerships were announced with companies like Nexuni, Foxconn, and Siemens for testing applications in laundry, factory automation, and industrial settings, while Isaac Lab 3.0 reported a 2.5x speedup over the previous version on visual dexterity benchmarks.
Source: Chinese robotics — Company news discovery · Read original article ↗
Article text · Machine translation into English
On this page
News
NVIDIA launches GR00T 2 and Isaac Lab 3.0 at CoRL 2026, driving robot training through simulation tools, world action models, and agentic robot workflows.
realtime news November 10, 2026 04:36 UTC
NVIDIA revealed breakthrough progress in robotics technology at the 2026 Robotics Learning Conference in Austin, Texas. The company launched GR00T 2, a world action model, and Isaac Lab 3.0, an open-source robot learning framework, to improve the precision and scalability of robot training. These innovations aim to bridge the gap between laboratory demonstrations and real-world applications by enabling robots to perform complex and dexterous tasks with greater autonomy and adaptability.
GR00T 2: Leap in World Action Models
GR00T 2 is based on NVIDIA Cosmos, aiming to train robots for dexterous manipulation using diverse datasets, including 100,000 hours of open-source human egocentric data. The main features of GR00T 2 include:
- Out-of-the-box performance: Successfully completed 50 assessment tasks out of 40 project tasks without specific task demonstrations.
- Skill Transfer: Enabling robots to learn from human demonstrations and adapt skills across different robot bodies.
- Generalization with minimal data: In 50 tasks, with one demonstration per task, a success rate of 71% was achieved, while five demonstrations increased it to 95%.
- Agent Control: Provides AI agents with tools to shape skill sequences and recover from errors.
GR00T 2 ranks in the top three in benchmarks such as RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces. Companies such as Nexuni, Foxconn, and Siemens are testing the model for robot laundry, factory automation, and industrial applications, respectively. MicroAGI also utilizes GR00T 2 for Ethernet cable swapping tasks via YAM robotic arms.
Isaac Lab 3.0: Scaling robot training in simulation
Isaac Lab 3.0 provides a powerful platform to train robots in simulated environments and then deploy them in the real world. The latest version introduces:
- Enhanced physics integration: Using the Newton physics engine to support deformable objects and complex joint systems.
- Increased training throughput: In vision-based dexterous manipulation benchmarks, it achieved 2.5x the training speed of the previous Isaac Lab 2.3.
- Flexible simulation tools: Allowing researchers to switch between NVIDIA PhysX and Newton solvers, choose renderers for perception-based training, and apply the reinforced post-training library to foundation models such as GR00T.
Skild AI announced the STEEL solver, a rigid-body physics engine built on Newton, to further accelerate robot training with Isaac Lab. Flexion demonstrated how Isaac Lab and Newton support its Reflect model, combining reasoning and whole-body manipulation.
Collaborative efforts and broader impacts
NVIDIA is collaborating with partners such as Ambi Robotics, Bosch Research, Siemens, and UC Berkeley to bring these advancements to the robotics community. The company has also launched new workflows, including Video to Data, which converts human videos into simulated robot demonstrations, and EgoHand, a hand motion reconstruction model with over 40,000 downloads, released on Hugging Face within 2 weeks.
By combining simulation, world action models, and agentic robot workflows, NVIDIA's latest technologies aim to accelerate the development of robots capable of performing complex tasks in dynamic environments, paving the way for broader industrial applications.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Data origin
- Human egocentric dataOpen source S6
Reported numbers
training_hours
100,000 hours
View original evidence
GR00T 2：世界动作模型的飞跃 GR00T 2 基于 NVIDIA Cosmos，旨在利用多样化的数据集训练机器人进行灵巧操作，包括 100,000 小时的开源人类自我中心数据。
Open source S6downloads
>40,000
View original evidence
该公司还推出了新的工作流程，包括将人类视频转换为模拟机器人演示的 Video to Data，以及具有超过 40,000 次下载次数的 EgoHand，这是一个手部动作重建模型，在 2 周内在 Hugging Face 上发布。
Open source S21training_speed_multiplier
2.5
task_count
40
task_count
50
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
GR00T 2：世界动作模型的飞跃 GR00T 2 基于 NVIDIA Cosmos，旨在利用多样化的数据集训练机器人进行灵巧操作，包括 100,000 小时的开源人类自我中心数据。
Open source S6
最小数据的泛化： 在 50 个任务中，以每个任务一个演示达到了 71% 的成功率，五个演示则提升至 95%。
Open source S9
该公司还推出了新的工作流程，包括将人类视频转换为模拟机器人演示的 Video to Data，以及具有超过 40,000 次下载次数的 EgoHand，这是一个手部动作重建模型，在 2 周内在 Hugging Face 上发布。
Open source S21
提高的训练吞吐量： 在基于视觉的灵巧操作基准中，达到了前任 Isaac Lab 2.3 的 2.5x 的训练速度。
Open source S16
GR00T 2 的主要特点包括： 即用型性能： 成功完成了 40 项目任务中的 50 考核任务，而无需特定任务演示。
Open source S7
Implications for data suppliers
RoboSignal interpretation and collection questions, not statements of buyer demand.
- Check whether the intended buyer needs human observations, hand or object annotations, or aligned robot actions. Human video alone does not provide robot commands.
- Compare the reported units and scope before using these quantities in a budget. Recording hours, sensor-hours and trajectories are different measures.
- Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.
Source：Chinese robotics — Company news discovery · blockchain.news