Gemini Robotics 2 brings whole body intelligence to robots

TL;DR

Google DeepMind’s direct successor to Gemini Robotics 1.5 ships three coordinated models: Gemini Robotics 2 (a vision-language-action model claimed to control full humanoids “from feet to fingertips” plus bi-arm robots), Gemini Robotics ER 2 (an upgraded embodied-reasoning VLM that acts as the high-level agent — plans multi-step tasks lasting minutes, calls tools, and coordinates the VLA), and Gemini Robotics On-Device 2 (an efficient VLA that adapts to a new bi-arm embodiment in a few hours with typically <200 examples). Headline capability moves over 1.5 are whole-body humanoid control on Apptronik’s Apollo 2, dexterous manipulation on a 22-DoF five-fingered SharpaWave hand (tying knots, sealing ziplock bags), longer-horizon agentic execution (“several minutes and involving hundreds of decisions”), and — new to the series — multi-robot collaboration where multiple robots coordinate on a workflow no single robot could finish. ER 2 is exposed on Google AI Studio and Gemini Enterprise Agent Platform; the VLA and On-Device models are partner-gated. Safety is packaged as ASIMOV-Agentic, a new benchmark for agentic safety orchestration (unsafe-tool refusal, task-possibility prediction, proactive human-intervention requests).

Key claims

Gemini Robotics 2 controls entire humanoid robots (not just the upper body) — walking, crouching, stretching, and manipulating in one intent-to-action loop; illustrated on Apptronik Apollo 2 executing a “put the watering can into the green bin in the bottom shelf” instruction that requires locomoting to and between two stations [§ “Humanoids in motion”].
Dexterity claim on a five-fingered 22-DoF SharpaWave hand: tying knots and sealing a ziplock bag; on parallel grippers (Franka Duo platform): complex tight-packing [§ “Bringing advanced dexterity to hands and grippers”].
Gemini Robotics ER 2 now handles task sequences “lasting several minutes and involving hundreds of decisions,” understands task begin/end boundaries, and pinpoints when key events occur — framed as a step change in progress understanding [§ “Unlocking advanced tasks with agentic reasoning and multi-robot collaboration”].
Multi-robot collaboration is introduced: different robot types can communicate and cooperatively execute workflows a single robot could not complete alone [§ “Unlocking advanced tasks with agentic reasoning and multi-robot collaboration”].
On-Device 2 inherits the “motion transfer” techniques from Gemini Robotics 1.5 and adapts to a new bi-arm embodiment in a few hours, typically with fewer than 200 examples, across drastically different shapes/sensors/DoFs — instantiated on Dexmate, SO-101, and Trossen platforms [§ “Adapting fast on-device models for any robot”].
New agentic-safety benchmark ASIMOV-Agentic measures: (a) the ER agent’s ability to refuse unsafe tool calls from a VLA, (b) prediction of task feasibility, and (c) proactive human-intervention requests under uncertainty [§ “Advancing our commitment to safe and responsible robotics”, ASIMOV-Agentic dataset card].
Gemini Robotics ER 2 is claimed as DeepMind’s safest robotics model to date on safety-constraint-following and human-proximity benchmarks — can detect nearby humans, trigger safety tool calls, and bring the robot to a safe stop when approached too closely [§ “Advancing our commitment to safe and responsible robotics”, Safety Technical Report].
Release posture: ER 2 available on Google AI Studio and (private preview) on the Gemini Enterprise Agent Platform; the VLA and On-Device VLA are restricted to early-access partners [§ intro, § release].

Method

The system continues the two-tier architectural pattern established in Gemini Robotics 1.5: an embodied-reasoning VLM (ER 2) acts as the high-level brain and a vision-language-action model (Robotics 2) executes low-level motor control, with ER 2 emitting per-step natural-language sub-instructions to Robotics 2 and monitoring progress. The new pieces in this update are: (1) the VLA’s output space extends to whole-body humanoid control (feet through fingertips) and to a 22-DoF five-fingered hand, not just upper-body / parallel grippers as in 1.5; (2) ER 2 has been trained/tuned to sustain execution over “several minutes and hundreds of decisions” with explicit begin/end-of-task and key-event detection heads used for progress tracking; (3) a new inter-robot coordination channel lets ER 2 fan work across multiple heterogeneous robots; (4) On-Device 2 is a smaller VLA optimized for local execution that inherits the 1.5 “motion transfer” adaptation recipe. No parameter counts, training-data sizes, or architectural details are given in the blog post; the accompanying Safety Technical Report PDF is the canonical source for the safety evaluations.

Results

The blog post carries no headline benchmark numbers. The substantive results asserted are: (a) whole-body humanoid execution of the watering-can-to-bin task on Apollo 2 (qualitative video); (b) knot-tying and ziplock-sealing on the SharpaWave hand and tight-packing on Franka Duo (qualitative); (c) On-Device 2 adaptation to Dexmate, SO-101, and Trossen with a few hours of data / <200 examples on new embodiments (qualitative); (d) safety improvements framed as SOTA on constraint-following and human-proximity benchmarks, with ASIMOV-Agentic proposed as a new benchmark rather than a scoreboard. Per-embodiment success rates, safety benchmark numbers, ablations, and comparisons against 1.5 or third-party VLAs (π0.5, XR-1, Spirit v1.5) are deferred to the Safety Technical Report and to yet-to-be-released write-ups; none are reproduced in the blog post.

Why it’s interesting

Gemini Robotics 2 is the direct successor to Gemini Robotics 1.5 brings AI agents into the physical world and updates the same two-tier ER + VLA recipe on three axes: expanding the VLA’s action space from upper-body to full whole-body humanoid + five-fingered dexterity, extending ER’s horizon from single-task planning to multi-minute / multi-robot orchestration, and shipping On-Device 2 as the fast-adaptation edge counterpart. It sharpens two active threads on the VLA Models concept page: (i) the closed-flagship VLA (Gemini Robotics 2 with a five-fingered dexterous hand + whole-body locomotion) is now the strongest filed embodied instance of the “single unified stack” position that the mimic + BFL announcement in FLUX 3 x mimic: The Next Generation of Video-Action Models targets from the open side; (ii) the multi-robot-collaboration claim moves the “orchestrator over interruptible tools” pattern tracked via RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning and Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents from single-robot orchestration to cross-robot orchestration — a fanout dimension no filed embodied paper covers yet. On Tool-Use Agents, ASIMOV-Agentic’s three axes (unsafe-tool-call refusal, task-feasibility prediction, proactive intervention requests) give the first filed benchmark that measures the safety of an embodied ER agent’s tool-use decisions, complementing but not overlapping with the sequential/parallel/critical-steps axes already tracked on that page.