Version 3

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 release covers whole-body humanoid control, on-device adaptation from under 200 examples, and multi-robot collaboration across heterogeneous platforms [2]. DeepMind also released an accompanying safety tech report and the ASIMOV-Agentic benchmark, now at version 2, designed to test whether orchestrating models can refuse unsafe robot commands [7][4]. Black Forest Labs took a different approach with FLUX 3, a single model trained jointly on images, video, audio, and robotic actions, with robotics as a downstream application of generative media training rather than a separate stack [9].

Physical AI is moving from demos toward deployable systems, but no fully independent third-party benchmark has yet verified zero-shot transfer of Gemini Robotics 2 to arbitrary platforms [10], leaving capability claims primarily supported by DeepMind's own evaluations. Whether ASIMOV-Agentic becomes an industry-shared safety standard or remains a self-referential benchmark will determine whether robot safety has common ground across labs.

Google DeepMind released Gemini Robotics 2.0 in late July 2026, framing 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 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 [3]. Of the three, only ER 2 is publicly available; the VLA and On-Device models remain limited to early-access partners [2].

DeepMind paired the release with a formal safety framework. The ASIMOV-Agentic benchmark, now at version 2, 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 [4][5][6]. DeepMind published an official safety tech report alongside the release [7], and researcher Anirudha Majumdar has publicly promoted the benchmark as covering unsafe task refusal, uncertainty quantification, and human escalation [8]. The benchmark is backed by peer-reviewed work in MLRS proceedings and on arXiv [5][6]. A notable capability across the stack is multi-robot collaboration: heterogeneous robots communicate via shared semantic understanding to hand off subtasks no single robot could complete alone [3][2].

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 [9]. The model was demonstrated on Audi production-line tasks involving cables, seals, and deformable parts that rule-based automation handles poorly [9]. Black Forest Labs reported 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 [9].

One gap runs across both systems: no fully independent third-party benchmark has yet shown reliable zero-shot transfer of Gemini Robotics 2 to arbitrary platforms [10], and FLUX 3's performance comparisons rest on self-reported internal evaluations. DeepMind's structured benchmark results are backed by peer-reviewed papers, but the most capable physical control models — VLA and On-Device 2 — remain behind early-access restrictions, limiting external scrutiny [2][1]. The question of whether any lab's safety or capability claims will be independently stress-tested is unresolved.