Gemini Robotics-ER: Embodied Reasoning

Google DeepMind's Gemini Robotics-ER 1.6 is not a new robot you can buy. It is a model for the part of a robot that has to look at the world, understand the task, check physical constraints, and decide whether the next action is safe or finished.

That layer is called embodied reasoning. For a home robot buyer, the plain-English version is: can the robot think about the room before it moves, and can it tell when its own work failed?

That matters because the next wave of home robots is already moving beyond voice assistants on wheels. 1X NEO, Samsung Ballie, Unitree G1, and future humanoids from Figure, Apptronik, and others all need some version of this loop before they can be trusted with chores. A robot that can start a task is useful. A robot that can notice the cup is about to tip, the drawer is stuck, or the room changed since yesterday is a different category.

What does embodied reasoning mean for home robots?

Most buyer-facing robot AI terms are overloaded. A VLA model connects vision, language, and action: it turns "pick up that mug" into motor behavior. A world model predicts what might happen next. Embodied reasoning sits above and around those systems. It asks practical questions about the physical scene:

• Which object is the user referring to?
• Is the requested object reachable from this body position?
• Is the item small, light, dry, and safe enough for this gripper?
• Did the robot actually complete the task, or did it only move in the right direction?
• Should the robot retry, ask for help, or refuse?

Google describes Gemini Robotics-ER 1.6 as a reasoning-first model for visual and spatial understanding, task planning, and success detection. It can call tools, including VLA models, rather than directly being the low-level motor controller. That distinction is important. In a home robot, you do not want one big model blindly issuing movement commands. You want a stack: perception, reasoning, action, verification, and safety limits.

The most useful way to think about embodied reasoning is as a chore supervisor. It may not be the hand that grasps the cup, but it should be the system that says, "the cup is too close to the edge," "the handle moved out of view," or "that request violates the robot's payload limit."

Gemini Robotics-ER 1.6: what actually changed

Google's announcement highlights four capabilities that map directly to home-robot problems.

Google says the 1.6 model improves spatial and physical reasoning over earlier Gemini Robotics-ER models and general Gemini models, especially on pointing, counting, success detection, and safety instruction following. The examples are industrial, but the failure modes are familiar in a kitchen or living room. Poor lighting, occlusion, ambiguous commands, and awkward camera angles are not factory-only problems.

The most buyer-relevant feature is success detection. ui44 has covered how robots know chores are done, because this is where many demos quietly cheat. Moving toward a target is easy to film. Knowing that the object landed in the right place, stayed upright, and did not create a new mess is the hard part.

Boston Dynamics is already using this commercially with Spot. Spot is not a home robot, but it is a useful preview because it works in messy real facilities rather than staged demo rooms. In the ui44 database, Spot is a 33.8 kg quadruped with roughly 90 minutes of battery life, autonomous navigation, self-charging, IP54 weather resistance, an optional arm, and a 14 kg payload capacity. Boston Dynamics says more than 1,500 units are deployed worldwide.

That is why the Google partnership is worth watching. Boston Dynamics says Gemini-powered AIVI-Learning can read gauges, measure sight-glass fullness, count pallets, detect puddles, and show transparent reasoning for inspection prompts. Those are not glamorous humanoid chores. They are the kind of boring verification tasks that make autonomy real.

Why homes are harder than Google’s examples

A pressure gauge is not a sock drawer. In some ways, it is easier: it is fixed to a wall, it has one purpose, and the success condition is a number. A home robot has to deal with soft objects, moving people, pets, clutter, sentimental items, and commands that are often incomplete.

That does not make Gemini Robotics-ER irrelevant to homes. It shows the missing middle layer. A home robot needs to reason about physical context before it calls the manipulation policy. For example:

• "Put the blue cup in the dishwasher" requires identifying the cup, locating the dishwasher rack, checking whether the cup is dishwasher-safe, avoiding a nearby knife, and confirming the cup did not fall sideways.
• "Tidy the living room" requires sorting toys, trash, remotes, clothes, cables, and personal items into different categories, then knowing when to stop.
• "Bring me my medication" requires reading labels, resolving ambiguity, refusing unsafe substitutions, and possibly asking for confirmation.

This is why flashy locomotion specs do not automatically translate into household usefulness. Unitree G1 is available from $13,500, stands 132 cm tall, weighs 35 kg, has a stated battery life of about two hours, and is offered in different configurations. Unitree’s specification table lists Dex3-1 three-finger hands as an option for G1 EDU, not the standard G1 associated with the $13,500 starting price. Those are impressive specs for a compact humanoid research platform. They do not tell you whether the robot can judge a messy bedside table safely.

The same applies to more home-focused robots. 1X NEO is listed in the ui44 database at $20,000 for early adopters, with a 167 cm soft humanoid body, 30 kg weight, roughly four hours of battery life, RGB and depth cameras, tactile skin, and a home-chore mission. Those physical choices make sense for homes. But the buyer question is still not "does it have cameras?" It is "can it use those cameras to avoid unsafe or pointless actions?"

The home-robot stack buyers should ask about

When a company says its robot has advanced AI, ask which layer it is talking about. A credible home robot needs at least five layers working together.

This is where Gemini Robotics-ER is more useful as a benchmark than as a brand name. If a robot maker cannot explain how its system checks success and handles constraints, it is probably still closer to a demo pipeline than a reliable home assistant.

Figure 03 is a good example of why the distinction matters. ui44's database lists Figure 03 as a 173 cm, 61 kg humanoid with a roughly five-hour battery, stereo vision, depth cameras, force sensors, tactile arrays, and a 20 kg payload. Figure's current public site is explicitly home-oriented: it calls Figure 03 "the future of home help," says it is built for everyday household tasks, and says Helix helps it navigate changing home environments. The buyer caveat is availability, not intent. Figure 03 still has no public consumer price or order path, while BMW and production-evaluation work remain separate evidence of where Figure can learn from controlled deployments.

Apptronik Apollo points in a similar direction. Apollo is a 173 cm, 73 kg enterprise humanoid with roughly four hours of battery life, vision, force/torque sensing, proprioception, and a reported heavy payload around 25 kg. Apptronik has a strategic partnership with Google DeepMind, and its official Apollo page says the robot will operate in warehouses and manufacturing plants in the near term before eventually extending into construction, oil and gas, electronics production, retail, home delivery, elder care, and more.

That production-learning path may be frustrating if you want a robot at home now, but it is rational. Factories and warehouses make it easier to define tasks, instrument outcomes, collect failure data, and roll out updates. Homes will inherit the useful parts only after the models prove they can reason outside the marketing video.

What embodied reasoning does not solve

Embodied reasoning is not magic. It will not make weak hardware strong, turn a toy arm into a dishwasher loader, or remove the need for safety certification. It also raises privacy questions because better reasoning often wants more camera views, more logs, and more data sharing.

Boston Dynamics is explicit that AIVI-Learning requires data sharing and that customer data is shared with Boston Dynamics to improve facility-specific models. That may be acceptable in an industrial inspection deployment. In a home, the bar is higher. Buyers should ask whether camera images leave the device, how long logs are stored, who can review failures, and whether the robot can keep useful functions when cloud AI is unavailable.

There is also a latency problem. Some reasoning can happen in the cloud. Some must happen locally. If a robot is holding a glass, stepping around a child, or deciding whether a drawer is jammed, it cannot wait politely for a remote model every time. The long-term answer is probably a split stack: local safety reflexes and motion control, plus larger off-board reasoning for slower planning and learning.

Samsung Ballie shows the consumer version of this tension. The ui44 database lists Ballie as an unreleased AI companion concept with announced Gemini plus Samsung language models, SmartThings integration, a camera, spatial and environmental sensors, projection features, and no confirmed price or release date. A rolling companion robot may not need humanoid manipulation, but it still needs embodied reasoning to know where it is, what it is looking at, whether the user’s request makes sense, and when not to disturb the household.

How to evaluate embodied reasoning claims before buying

For the next few years, most embodied-reasoning claims will sound better than the products feel. Use this checklist before putting down a deposit or joining a beta.

1. Ask for failure handling, not only success clips

A credible demo should show what happens when the robot misses a grasp, sees two similar objects, finds a blocked path, or fails to complete the task. Does it retry intelligently? Does it ask for clarification? Does it stop safely?

2. Look for multi-view verification

A wrist camera alone can lose the scene. A room camera alone can miss the hand. Multi-view reasoning matters because chores often fail in small, hidden ways: a cable drags behind a basket, a cup tips inside a rack, or a folded shirt slips off the pile.

3. Separate navigation autonomy from chore autonomy

A robot that maps your house is not automatically a robot that can perform chores. Navigation is necessary. Manipulation plus success detection is the leap.

4. Check local safety limits

Ask what the robot refuses to handle: liquids, knives, pets, medicines, glass, hot objects, heavy items, or stairs. A useful home robot should have clear boundaries. ui44's home robot safety guide covers why saying no is a feature, not a defect.

5. Ask what improves after purchase

If the company says the robot learns, ask from what data. Does it learn only from your home? From fleet failures? From teleoperation corrections? From simulation? From app updates? And can you opt out?

Build an evaluation around an uncertain result

A useful acceptance discussion starts with an ordinary, supported task rather than an adversarial stunt. Consider asking a future household robot to move a clean, empty plastic cup to a designated tray. This is an illustrative evaluation plan, not a claim that Gemini or any robot named here has passed it. The supplier would first need to confirm that the object, location and action fall within its operating instructions.

Define the finish condition before watching the attempt: the correct cup is on the correct tray, remains upright and has been released. Record the initial arrangement and any preparation by the demonstrator. Then ask how the system would distinguish three outcomes: verified completion, visible failure and an outcome it cannot see well enough to judge. That third category matters. A robot should not earn a successful result merely because its camera lost sight of the cup at the moment of release.

Google's ER 1.6 announcement identifies success detection and multi-view understanding as capabilities. It does not make an installation-specific promise about this cup task. The practical question is which evidence the implemented system uses: a final image, a second camera, gripper feedback or another observation. Ask the supplier to explain what happens when these observations disagree. A readable explanation of success is useful for troubleshooting; it is not itself an independent measurement that the task succeeded.

For a comparable record, retain every attempt in the agreed session, including aborts and interventions. Put human setup time beside robot execution time. If an operator identifies the cup or corrects the destination, label that help rather than hiding it inside the result. Keep the hardware, software version and task definition attached to the record. Results from one well-lit tray do not establish performance with reflective glass, stacked dishes or a cluttered room. Any extension needs its own evidence within the supplier's supported limits.

Ask how a changed model becomes a supported product

Boston Dynamics' AIVI-Learning announcement describes cloud model updates and required data sharing. Those are statements about its industrial service, not terms for a future household robot. They do, however, expose an ownership question that a hardware specification misses: which party is responsible when the reasoning service changes?

Ask the robot supplier how it checks a model update against the tasks it sells, what owners can learn about changed behavior, and what support does when an established task stops working. A household should not have to diagnose whether the camera, network, action controller or remote model caused the failure before it can request help. The contract should identify a support contact and the information needed to investigate a reported problem.

Also request the documented behavior when reasoning is unavailable. Losing an internet connection before a task begins differs from losing it during a task. The relevant answer is the product's specified fallback, including who must be present and what assistance may be required. Do not test outages during motion unless the manufacturer explicitly provides an appropriate procedure.

Finally, distinguish data needed to operate the service from data used to improve it. Ask whether those purposes have separate choices and whether declining an optional use changes the functions you intend to buy. The purchase decision should reflect the actual service conditions, not an assumption that every robot using a familiar model inherits the same privacy controls or update policy.

What embodied reasoning changes for buyers

Gemini Robotics-ER 1.6 will not put a humanoid in your kitchen tomorrow. But it clarifies what the credible companies are racing to build: not just robot bodies, and not just chatty AI, but a reasoning layer that can connect perception, action, success detection, and safety.

For buyers, that changes the question. Do not ask only whether a robot has a VLA model, a humanoid body, or a famous AI partner. Ask whether it can understand the physical consequences of its own actions.

The first genuinely useful home robots may not be the ones with the most human-looking hands or the fastest walking speed. They may be the ones that can pause, check the scene, notice the chore is not done, and choose the safer next move.