Google DeepMind's Gemini Robotics 2.5 hits 96% sorting success on Mercedes production line
量子位 · wechat · 2026-11-11
Google DeepMind unveiled Gemini Robotics 2.5, its first robot foundation model accelerated by Gemini 4 Argon with recursive self-improvement (RSI) capabilities.
- Capabilities: across 200 tasks, success rates improved 30–100% over the previous baseline with 20–30% faster completion. Dexterous soft-object handling jumped: cake-picking went from 5% to 100%, untying bows from 0% to 70%, hanging cardigans to 62%, folding pants to 40%.
- Factory work: in proof-of-concept tests at Mercedes plants in Sindelfingen and Kecskemét, Apptronik's Apollo 2 humanoid (22-DOF five-finger hands) sorted car parts at a 96% success rate, with the ER variant planning and 2.5 executing.
- RSI pipeline: Gemini 4 Argon can reconstruct physics simulation environments from a single RGB photo, auto-generate varied practice scenarios, and analyze failure videos to propose fixes — an AI coach accelerating robot development end to end.
- Safety: inference-time safety guidance improved obstacle-avoidance task performance by 33% and near-human operation by 43%; on the new ASIMOV-Jailbreak benchmark, Argon achieved a 99.5% jailbreak resistance rate vs. 82% for the prior baseline.
Google demonstrates a concrete path to "using stronger AI to train the next generation of robot AI," though fully autonomous closed-loop RSI remains distant.
More from Embodied
- NVIDIA's VLA driving model hits 20 decisions/s on Jetson Thor after Embedl optimization — No-Dragonfly6246 · 2026-11-13
- UC Berkeley's ATLAS drives humanoid robots from a single human video, no VLA or task-specific training — ChongZzZhang · 2026-11-13
- Roboticist: You Can't Do Robotics Without NVIDIA's Stack; Jensen's Long Bets Are Underrated — tianyin_xu · 2026-11-13
- US humanoids span pilots, early commercial expansions and production ramps amid FCC import limits — ingliguori · 2026-11-13
- Invictus Robotics unveils Aeras, a force-controlled aerial manipulation drone for industrial work — lukas_m_ziegler · 2026-11-13
- Zero-HLM: 45 sim demos become 4,800, humanoid zero-shot sim2real hits 69% vs 23% baseline — 机器之心 · 2026-11-13