Version 2

Google DeepMind released Gemini Robotics 2.0 in late July 2026, comprising three sub-models: Gemini Robotics ER 2 (publicly available via the Gemini Live API), a vision-language-action model, and an on-device model for hardware-constrained deployment [2][1]. The system provides whole-body humanoid control, on-device adaptation from under 200 examples, and multi-robot collaboration across heterogeneous platforms [2]. Separately, Black Forest Labs launched FLUX 3 around July 25 — a single foundation model trained jointly on images, video, audio, and robotic actions — with robotics demonstrated on Audi factory tasks [7]. DeepMind also released the ASIMOV-Agentic benchmark, backed by peer-reviewed papers and an official site, to test whether foundation model orchestrators can refuse unsafe robot commands [4][5][6].

Physical AI is moving from demos toward deployable systems, with DeepMind claiming production-grade latency via a public API and on-device operation without network connectivity. Whether the ASIMOV-Agentic benchmark gains adoption beyond DeepMind's own evaluations will determine whether there is a shared standard for robot safety — or whether each lab defines safety on its own terms.

Google DeepMind released Gemini Robotics 2.0 in late July 2026, describing its goal as building a 'generalist robot' capable of executing arbitrary human-directed tasks — what its scientists call 'physical AGI' [1]. The release covers three interconnected models. Gemini Robotics 2 provides whole-body humanoid control from feet to fingertips, enabling robots to walk, crouch, and manipulate objects in unstructured environments [2][1]. Gemini Robotics On-Device 2 is designed for hardware-constrained deployment: it adapts to new robot morphologies in a few hours using fewer than 200 demonstration examples, running locally without network access [2]. Gemini Robotics ER 2, the embodied reasoning layer, achieves 91.3% accuracy on moment-finding tasks at 4x the speed of competing larger models, and integrates with the Gemini Live API for sub-second latency in real-time task orchestration [3]. Of the three, only ER 2 is publicly available; the VLA and On-Device models remain limited to early-access partners [2].

A notable feature across the stack is multi-robot collaboration: heterogeneous robots — humanoids alongside bi-arm platforms — communicate via shared semantic understanding to hand off subtasks no single robot could complete alone [3][2]. On safety, DeepMind introduced the ASIMOV-Agentic benchmark, which tests whether high-level foundation model orchestrators can refuse unsafe tool calls from lower-level action models and seek human clarification when physical feasibility is uncertain [3][2]. The benchmark is backed by peer-reviewed work in MLRS proceedings and on arXiv, with an official site at asimov-benchmark.github.io [4][5][6].

Black Forest Labs took a different architectural approach with FLUX 3, launched around July 25. Rather than a specialized robotics stack, FLUX 3 is a single model trained jointly on images, video, audio, and robotic actions, with robotics treated as a downstream application of generative media training rather than a separate engineering effort [7]. The model was demonstrated on Audi production-line tasks involving cables, seals, and deformable parts that rule-based automation handles poorly [7]. Black Forest Labs reported that FLUX 3 outperformed Runway Gen-4.5 in 77% of comparisons and Luma Ray 3.2 in 93%, though these figures come from internal evaluations [7]. The underlying argument is that generating convincing video of physical tasks requires internalizing causal rules about how the physical world behaves, making generative video and robot action training mutually reinforcing.