Google DeepMind used July 30, 2026 to push Gemini Robotics 2 into the spotlight as a vision-language-action stack for whole-body humanoid control, feet to fingertips, demonstrated on Apptronik's Apollo 2. That sentence alone would have been science fiction marketing five years ago. Today it arrives with sibling pieces that matter almost as much as the walking demos. ER 2 is the embodied reasoning line surfaced through AI Studio. On-Device 2 is the edge-oriented sibling. The main VLA path is private preview for trusted testers, not a download button for every robotics startup with a GPU budget. DeepMind also pointed at ASIMOV-Agentic as a safety benchmark in the same wave. The videos are polished. The dexterity numbers are not. Trash bag tasks around 44 percent, ziplock around 40 percent, unscrewing a bulb around 92 percent, dustpan work around 32 percent. Walking, crouching, and shelf tasks show up alongside multi-robot collaboration. I watched the package as a reporter who has been burned by robot GIFs before. The capability is real enough to cover. The success rates are the adult supervision.
Whole-body VLA on Apollo 2, and what that phrase means
Whole-body control is the claim that separates this from arm-only tabletop demos. Feet to fingertips means balance, locomotion, and manipulation live in one policy story instead of a walking stack glued to a separate gripper brain. Apptronik's Apollo 2 is the humanoid body DeepMind used for the showcase. That pairing matters because hardware choice shapes what failure looks like on camera.
Vision-language-action framing puts Gemini Robotics 2 in the same broad family as other VLA efforts. Perception and language conditioning feed actions. The July 30 materials present that loop as competent enough for walk, crouch, and shelf work, plus collaborative behaviors across more than one robot. Collaboration is easy to oversell in a video edit. It is still a different product ambition than a single arm folding shirts in a guarded cell.
Private preview and trusted tester access is the distribution reality check. If you cannot pull the weights or book an API without a relationship, you are watching a controlled release. That is normal for embodied models with safety and hardware coupling. It also means most of the industry will evaluate marketing artifacts longer than they evaluate the model on their own floor.
ER 2 through AI Studio is the more reachable surface for people who live in software. Embodied reasoning without a full humanoid on your loading dock is how DeepMind can widen the audience beyond labs with Apollo access. On-Device 2 speaks to latency, privacy, and untethered operation themes that every robot integrator eventually hits when the WAN link drops.
I am treating the three names as a portfolio. Gemini Robotics 2 VLA is the whole-body humanoid tip of the spear. ER 2 is the reasoning layer you can poke in AI Studio. On-Device 2 is the deployment-shaped sibling. Confusing them into one blob makes the launch look bigger and less precise than it is.
ASIMOV-Agentic appearing as a safety benchmark is DeepMind signaling that agentic robot behavior needs its own exams. I welcome that more than another cherry-picked hallway walk. Benchmarks can be gamed. Their absence is worse.
The dexterity table is the most important slide
Trash bag about 44 percent. Ziplock about 40 percent. Unscrew a bulb about 92 percent. Dustpan about 32 percent. Those are not vibes. Those are the numbers that should travel with every embed of the demo reel. A bulb task near ceiling performance next to dustpan work near one in three tells you the skill distribution is uneven in the way real households and warehouses are uneven.
I trust uneven numbers more than a silent montage. If every chore were ninety-plus, I would assume a narrow stage set. Thirty-two percent on dustpan work is a confession that contact-rich, deformable, oddly shaped chores remain brutal. Forty percent on ziplock is the same confession with a kitchen prop. Publish that. Do not bury it under triumphant music.
Walking, crouching, and shelf tasks fill the locomotion and reach story. Those clips are what executives forward. They should forward them with the dexterity table attached. Locomotion progress without manipulation reliability is how you get a very expensive tour guide that still cannot clean up after itself.
Multi-robot collaboration is the ambition line. Coordinated behavior is where warehouses want to go. It is also where safety cases get harder, because failures compound across machines. ASIMOV-Agentic exists in this launch story for a reason. Agentic coordination without a safety yardstick is how demo days turn into incident reports.
Curated videos are a disclosure, not an insult. Every robotics lab curates. The obligation is to keep the success rates in the same browser tab as the MP4. DeepMind gave enough numeric texture here that lazy secondary coverage has no excuse to pretend the system is production-ready chore labor.
My working summary for operators is simple. Useful research prototype with whole-body reach on Apollo 2. Not a staffing plan. Not a replacement schedule for night-shift cleaning. Not a green light to fire the people who currently hold trash bags and dustpans.
Private preview, safety theater risks, and how I am filing this
Trusted tester access means the first external lessons will arrive filtered through partners who already speak DeepMind's hardware dialect. That can be good for safety. It can also slow adversarial evaluation from independent labs. I want both careful onboarding and eventual broader scrutiny.
ER 2 in AI Studio may become the thing developers actually touch first. If embodied reasoning APIs teach better planning for physical tasks even before you own a humanoid, that could matter more for near-term software than Apollo footage. On-Device 2 is the piece I will watch for factory networks that refuse cloud round trips.
The honest line, and DeepMind's own rates support it, is that demos are not production. Ninety-two percent on a bulb is encouraging. Thirty-two percent on a dustpan is a product veto for unsupervised deployment in messy spaces. Anyone blurring those into a single wow should be removed from your vendor evaluation committee.
Competitors will answer with their own humanoid reels within the news cycle. That race is now table stakes. The differentiator worth caring about is published failure structure, safety benchmarks like ASIMOV-Agentic, and whether private preview partners report transfer beyond the filmed set pieces.
I will file Gemini Robotics 2 as a meaningful whole-body VLA milestone on Apollo 2 with accompanying ER 2 and On-Device 2 threads, gated by trusted tester access, accompanied by dexterity numbers that still scream research. That is a strong week for embodied Gemini. It is not a strong week for believing robot chores are solved.
If you build on this stack later, demand the same numeric honesty in your own acceptance tests. Film the wins. Count the losses. Ship when dustpan-class tasks stop looking like coin flips. Until then, keep humans in the loop and keep the highlight reel out of the procurement contract.
- LLMs
