THE APEX TIMES
At the Conference on Robot Learning in Austin, NVIDIA and partners showcased how robots can rehearse delicate manipulation tasks in simulation, using open models, physics tooling, and new benchmarks to improve reliability outside the lab.
NVIDIA used the Conference on Robot Learning (CoRL) in Austin, Texas, to argue that the next jump in robotics autonomy will come less from teaching robots to walk and more from rehearsing complex, precise hand and tool work in simulation. The company’s message was that learning to perform delicate manipulation, then repeating that performance across changing real-world conditions, requires both better “world models” and more scalable ways to test policies before they touch hardware.
CoRL this week has put a spotlight on agentic robotics, a paradigm that combines frontier AI agents with the real-world systems that must execute physical actions. NVIDIA’s approach centers on what it calls world action models, which it describes as enabling robots to perceive their surroundings, anticipate how actions will change the environment, and then carry out multistep tasks with high precision. The company is framing these tools as a bridge from demonstrations in controlled settings to industrial deployment where small errors can cause failures.
Among the headline announcements is GR00T 2, an open frontier world action model built on NVIDIA Cosmos. NVIDIA said GR00T 2 was trained on a mix of robot and human data, including 100,000 hours of NVIDIA open source human egocentric data. The company linked GR00T 2 to dexterous manipulation capabilities, and it also reported performance standing on several named benchmarks. NVIDIA said GR00T 2 ranks among the top three models across vision language action and world action model categories for the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces benchmarks, citing multiple leaderboard wins.
NVIDIA pointed to early industrial testing of GR00T 2 by several partners. It said Nexuni is evaluating the model for robotic laundry handling, Foxconn is applying GR00T 2 to factory automation for server manufacturing, and Siemens is evaluating it across industrial use cases. It also described MicroAGI’s work on Ethernet cable swapping using a YAM robotic arm, where MicroAGI applies reinforcement learning to post-train GR00T 2 and uses an agentic robotics workflow to automate simulation environment creation and calibration.
On the robotics software side, NVIDIA emphasized Isaac Lab 3.0, an open source robot-learning framework it says helps researchers set up training environments faster using simulation-ready assets. NVIDIA said Isaac Lab 3.0 supports switching between two physics solvers, NVIDIA PhysX and Newton, and it also offers renderers for perception-based training. A key addition, according to NVIDIA, is a new reinforcement post-training library that can apply to foundation models such as GR00T.
NVIDIA’s physics and simulation accuracy is positioned as essential for “sim-to-real” transfer, where skills learned in simulation must hold up when deformable objects and contact-rich interactions show up. The company said tighter integration with Newton now allows Isaac Lab to model deformable materials such as cables and cloth, and to simulate complex articulated systems like robotic hands with many interacting joints. In one example, NVIDIA said Newton integration can simulate how a strawberry deforms under a robot’s grip.
NVIDIA also disclosed performance and scale claims tied to the Isaac Lab 3.0 release. It said Isaac Lab and Newton together enable researchers to train thousands of simulated robots in parallel and achieve up to 2.5x the training throughput in a vision-based dexterous manipulation benchmark, comparing Isaac Lab 3.0 with Newton against Isaac Lab 2.3 with PhysX. NVIDIA further described parallel development efforts around Newton, including a STEEL (Scalable Throughput-optimized Engine for Embodied Learning) physics solver announced at CoRL by Skild AI, designed to accelerate robot training with Isaac Lab.
Beyond training, NVIDIA highlighted evaluation and benchmark work aimed at measuring dexterity and adaptability under varied conditions. It described Isaac Lab-Arena, an open source framework for scaled policy evaluation that lets developers compose diverse, repeatable evaluations and run policies across thousands of variations in parallel. NVIDIA said a developer preview of the latest version, 0.4, is arriving at CoRL, and it said analyses can reveal which conditions cause failures to guide the next training run. New open benchmarks at CoRL include DexBench and VABAR. NVIDIA said DexBench, built using Isaac Lab-Arena, includes 15 tasks across four scenes covering retail, logistics, manufacturing, and hospitality, testing generalization across object, spatial, and physics variations. VABAR (Variational Automation Benchmark for Agentic Robotics), from UC Berkeley and NVIDIA, evaluates robots across six industrial tasks with standardized variations, difficulty tiers, and success metrics.
The company also addressed a remaining gap: creating reliable programs for multistep tasks when something goes wrong. NVIDIA pointed to UC Berkeley’s Graph-as-Policy (GaP) framework, which it said uses agents to connect individual robot skills into robot programs, then tests those programs in simulation and refines them based on failures. NVIDIA quoted Ken Goldberg, a UC Berkeley professor of robotics and a lead on VABAR and GaP, saying agentic robotics is among the most exciting technological developments since the internet. NVIDIA said partners including Ambi Robotics, Bosch Research, and Siemens are exploring GaP-style workflows for industrial scenarios.
A sharp selloff following Intel’s third-quarter results has drawn renewed attention to Intel’s cap-structure and the government’s role as a major shareholder, turning Wall Street frustration into a broader political question about public exposure to corporate performance.
As Google rolls out additional TPU options aimed at generative AI workloads, Nvidia is likely to be judged on whether its hardware, software stack, and supply can keep winning in data centers amid heightened leadership attention.
A CNN report circulated through markets on Monday, tying a probe involving Nvidia Chief Executive Jensen Huang to allegations of a personal relationship with an alleged Chinese spy. The company’s disclosure, timing, and the scope of any investigation were not addressed in the report.
After an earnings update, Advanced Micro Devices said growth is continuing and voiced optimism for the fourth quarter, even as market attention has shifted toward concerns that the so-called AI trade is cooling.
A new market commentary argues that the next wave of “autonomy” spending in defense may be flowing toward Shield AI, even as Palantir remains a prominent software name in military analytics.
A Yahoo Finance report says Jensen Huang’s personal ties could put Nvidia under greater scrutiny in an era where U.S. national-security concerns increasingly intersect with advanced chip supply chains.
A sharp selloff in Alphabet’s stock has erased nearly $500 billion in market value, reflecting investor skepticism about how quickly heavy AI spending will translate into profits.
A fresh note from Goldman Sachs is prompting investors to look past Nvidia’s recent momentum, after the company’s shares moved lower and analysts questioned what the market may be underpricing in the AI hardware boom.
A report attributed to an unspecified foreign intelligence effort has triggered a board-level investigation at Nvidia, adding a governance and geopolitical risk overlay to investor sentiment around the AI chipmaker.
Culper Research had raised concerns about unusually close contacts involving Nvidia CEO Jensen Huang months before the reported “spy scandal” became public, according to a Yahoo Finance account.
Comments
Community guidelines. Comments may be removed for violations.
Loading comments...