NVIDIA ships Isaac Lab 3.0, Isaac OS and a reference humanoid at GTC Berlin
AspireAI · NVIDIA Newsroom
NVIDIA announced at GTC Berlin that Isaac Lab 3.0 is now generally available: an open source, agent-ready framework for training and refining robot policies from a single desktop up to a data centre, using multi-GPU workflows and physics engines including Newton and NVIDIA PhysX. Alongside it came Isaac OS, a new open source runtime built on Holoscan that connects AI, sensors, controls and hardware for deterministic real-time performance, carrying policies from simulation onto physical robots.
A reference humanoid arrived with the release. Agile Robots announced the Agile ONE SN, a full-stack research platform preconfigured with NVIDIA Isaac GR00T and powered by Jetson Thor. Specs from the release: 1.40 m tall, 40 kg, 31 degrees of freedom, 3 kg payload per arm, dual WUJI Hand 2 five-finger hands with 20 degrees of freedom each, 1 kHz force control and tactile sensing, and Jetson Thor onboard compute at 2,070 FP4 teraflops with 128 GB of unified memory. NVIDIA says the unified platform cuts setup time from weeks to days.
Also announced: three open source reference workflows for DGX Station that cover reinforcement-learning refinement of GR00T 1.7 models, dexterous manipulation training in Isaac Lab, and post-training of Cosmos world models for synthetic data generation using ChatGPT Astra; a SimReady Central Library of more than 2,000 simulation-ready robotics and factory assets on Hugging Face; the Human 1M Dataset with synchronized three-view video and 3D hand tracking, whose full version is restricted to Cosmos Coalition partners; and the Halos safety platform built on IGX Thor.
Why it matters: the recurring bottleneck in robotics is not the model, it is the plumbing between simulation, training and deployment. This release attacks the plumbing, and the reference humanoid gives researchers a shared hardware target instead of everyone assembling their own.
Honest limitation: the weeks-to-days setup figure is NVIDIA's own characterization, not an independent measurement, and the release carries the standard caveat that many of the described products and features remain in various stages and are offered on a when-and-if-available basis.
Source: NVIDIA newsroom press release, October 21, 2026.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.