Gemini Robotics Embodied Reasoning 2 reached public preview on Google Cloud, putting a robot's high-level brain behind enterprise IAM, simulation and fleet telemetry
AspireAI community · Google Cloud Blog
Google Cloud announced on October 27, 2026 that Gemini Robotics Embodied Reasoning 2 - the specialized Gemini model that acts as a robot's high-level brain - is now in public preview on the Gemini platform via Model Garden. The model was introduced in July 2026 as a reasoning layer that works across any embodiment, from stationary arms to humanoids, combining spatial and temporal video understanding with long-horizon planning and built-in safety guardrails. Developers declare low-level Vision-Language-Action models or navigation APIs as tools and stream multimodal video, audio or text into it, so it can orchestrate multi-minute tasks involving hundreds of decisions. The Cloud release adds what a research demo does not have: enterprise identity and access management, multi-agent orchestration, simulation in MuJoCo-Warp and NVIDIA Isaac Sim, synthetic scenario generation with auto-labelled workcell imagery, Google Distributed Cloud for latency-sensitive local work like SLAM and path planning, and BigQuery fleet telemetry. In Google's own July evaluations the model reached 57.4% accuracy on five-level progress classification and 91.3% accuracy with a 0.96 second mean absolute distance on moment-finding. This is a preview, not a general release, and the numbers come from Google's own benchmarks.
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.