Feet to Fingertips: Gemini Robotics 2 Targets Whole-Body Control as Trade Policy Hits Robot Imports

Humanoid robots spent years as spectacle—falling onstage, juggling viral demos, rarely earning a purchase order. Google DeepMind's July 30, 2026 release of Gemini Robotics 2 argues the category is entering a different phase: whole-body control, multi-finger dexterity, and multi-robot coordination packaged as a deployable intelligence layer rather than a single stunt policy.

The announcement lands amid a separate policy shock: the U.S. Federal Communications Commission added foreign-produced advanced robotic devices—including humanoids and quadrupeds—to its Covered List, blocking new import authorizations on national-security grounds. Technology and trade are now intertwined for physical AI in ways that lab benchmarks alone do not capture.

This article focuses on what DeepMind actually shipped; the FCC rule is context, not the engineering core.

Three models, one stack

Gemini Robotics 2 is not a single checkpoint. DeepMind released a trio:

Gemini Robotics 2 (VLA): A vision-language-action model mapping vision and language to motor control. It controls full humanoids "from feet to fingertips" and bi-arm platforms, with improved dexterity on five-finger hands and parallel grippers.

Gemini Robotics ER 2 (embodied reasoning): A vision-language model acting as a high-level agent—planning multi-step tasks over several minutes, communicating with humans, and coordinating multiple robots. ER 2 is available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform.

Gemini Robotics On-Device 2: A lightweight VLA optimized for local execution without network latency. DeepMind says it adapts to new robot embodiments in "a few hours" with fewer than 200 examples—even when morphology, sensors, and degrees of freedom change sharply.

The architecture mirrors how production robot fleets will likely deploy: reasoning in the cloud or edge agent layer, control on-device where connectivity fails, VLA executing motion primitives.

Whole-body motion: beyond table-top reach

Prior Gemini Robotics releases emphasized upper-body manipulation at tables. Version 2 adds locomotion and posture: walk, crouch, stretch, and manipulate in cluttered spaces.

DeepMind demonstrated Apptronik's Apollo 2 humanoid following instructions such as placing a watering can into a green bin on a bottom shelf—walking to the table, grasping, repositioning, and placing with whole-body coordination.

The blog candidly notes robots "have more to advance in movement speed," framing whole-body control as a prerequisite for complex real-world tasks rather than a finished product.

Dexterity: trash bags, knots, and parallel grippers

Five-finger manipulation remains robotics' hardest showroom test. DeepMind reports Gemini Robotics 2 controlling the 22 degree-of-freedom SharpaWave hand on Apollo 2 for tasks including tying knots and sealing Ziploc bags. On a Franka Duo with Robotiq grippers, the same model checkpoint handles tight packing dexterity.

Success-rate charts in the release show medium-to-high performance on whole-body and gripper tasks, with multi-finger manipulation still listed as challenging—honest benchmarking rather than cherry-picked perfection.

Multi-robot collaboration and longer horizons

Gemini Robotics ER 2 upgrades task tracking: robots observe continuous video, detect when subtasks begin and end, self-correct on failure, and—new in this release—coordinate with other robots in shared spaces.

DeepMind positions ER 2 as the "brain" that plans and delegates execution to any lower-level VLA. That split matters commercially: hardware vendors can swap motor policies while retaining a common reasoning API.

Safety: ASIMOV-Agentic and proximity stops

Each robotics release from DeepMind has emphasized layered safety. Gemini Robotics 2 adds ASIMOV-Agentic, a benchmark for agentic safety orchestration measuring whether embodied reasoning agents refuse unsafe tool calls from VLAs, predict task feasibility, and request human help when uncertain.

DeepMind says Gemini Robotics ER 2 is its safest model yet on constraint-following and human-proximity benchmarks—detecting nearby people and triggering safe stops, aligning with collaborative robotics standards.

The ASIMOV-Agentic benchmark is published on Hugging Face in full, allowing third-party evaluation—a rarity in embodied AI where safety claims often outrun public test suites.

Availability and the FCC backdrop

VLA and On-Device models are early-access partner releases; ER 2 is the piece developers can probe immediately via Google AI Studio.

Meanwhile, MIT Technology Review reported August 3 that the FCC ban on new foreign advanced robot imports cites both cybersecurity surveillance risk and industrial policy protecting U.S. robotics supply chains. The Association for Advancing Automation told the publication that roughly 90% of recent U.S. university robotics papers relied on Unitree platforms—affordable Chinese hardware that may face import restrictions.

DeepMind's software advance does not resolve that hardware access question. Labs that depended on sub-$5,000 quadrupeds for data collection may find model improvements outpace their ability to buy platforms—a tension the robotics community is only beginning to price in.

Bottom line

Gemini Robotics 2 is a credible step toward generalist physical AI: one VLA checkpoint across humanoids and arms, ER 2 planning over minutes not seconds, on-device adaptation in hours not weeks, and public safety benchmarks to scrutinize.

It is not general-purpose labor yet. Speed, reliability, and cost of humanoid hardware still gate deployment. But the release marks the moment whole-body humanoid control and five-finger dexterity moved from research montages into a named product generation—with APIs developers can actually call.

### Sources

• Google DeepMind — Gemini Robotics 2 brings whole body intelligence to robots (July 30, 2026)
• The Verge — Google DeepMind's new AI model can control a robot's entire body (July 30, 2026)
• Ars Technica — Google reveals Gemini Robotics 2.0, promising improved dexterity and safety (July 2026)
• MIT Technology Review — Trump's AI protectionism has come for robotics (August 3, 2026)