Why Carolina Parada says useful robots can arrive before AGI
Parada’s September interview puts existing robots first. Follow the Spot example, Google’s earlier model roadmap and the customer evidence still needed.
A defined inspection workflow can be tested before general-purpose robotics is solved. The important evidence is whether added reasoning improves the customer’s actual work.
Carolina Parada is arguing for useful robots before robots that can do everything. In an Automated interview published September 17, Google DeepMind’s robotics chief described adding reasoning to machines already doing jobs. She pointed to Boston Dynamics’ Spot inspection system, while keeping general intelligence in the physical world as the longer-term ambition. The interview, from 32:54
The practical question is what the extra reasoning changes. If a robot already reaches the right place and captures an image, understanding that image more flexibly could expand the inspection it performs. That is a smaller, testable proposition than replacing an entire worker. It also gives customers a way to evaluate progress without agreeing on a definition of artificial general intelligence, or AGI.
September clarifies the route from research to use
Parada’s July 30 Gemini Robotics 2 announcement described three model roles: reasoning about a task, turning observations and instructions into robot actions, and running action models locally on the robot. The announcement also said that fine manipulation with multiple fingers remained challenging. It presented progress toward broader capability, with limitations still visible. The July announcement
The September conversation gives that research a commercial sequence: improve what existing machines can do while developing more general systems. Our reading is that early deployments can be evaluated on their own merits. A better inspection workflow would be a useful result even while a different robot struggles with an unfamiliar manipulation task.
Spot provides a specific case to check
Boston Dynamics’ own rollout account says it integrated Gemini and Gemini Robotics ER 1.6 into Orbit AIVI-Learning. It describes visual checks such as reading gauges, counting pallets and detecting standing liquid. The company says the update is live for enrolled AIVI-Learning customers. This is an existing integration discussed in September, not a new launch announced by the interview. Boston Dynamics’ account
There is an operational condition behind the feature list. Boston Dynamics says the service uses models updated in the cloud and requires data sharing. That should not be confused with the separate on-device action model in Google’s July release. A prospective deployment needs to establish which part runs where and what access to site data it requires.
| Question | Useful customer evidence |
|---|---|
| Does it read the asset correctly? | Accuracy under the site’s actual lighting, angles and equipment conditions. |
| Does the alert help? | Missed problems and unnecessary alerts, compared with the previous inspection method. |
| Does the workflow save effort? | Time spent reviewing results, resolving uncertainty and maintaining the system. |
Availability is the start of the business test
The named integration makes the discussion more concrete, but a product description cannot establish a customer’s savings. The next useful disclosure would compare the same inspection workload before and after the AI feature was added, over a stated operating period. That would help separate better model behavior from a measurable improvement in the job.
For readers following Google’s robotics strategy, repeated customer use and expansion would be stronger commercial signals than a longer list of supported checks. Persistent review work or unreliable results would weaken the case. The evidence to watch is the work accomplished and the cost of accomplishing it.
- Customer comparisons of inspection accuracy, false alerts and review time.
- Evidence of repeated use and expansion after initial feature access.
- Clear operating requirements for cloud access and sharing site data.
Sources & context
- Carolina Parada on Gemini Robotics and Google DeepMind’s Physical AI StrategyAssociation for Advancing Automation · Sep 17, 2026 · Primary sourceEpisode 58 transcript, 00:32:54–00:35:52 and 00:23–00:27
- Gemini Robotics 2 brings whole body intelligence to robotsGoogle DeepMind · Jul 30, 2026 · Primary sourceDated Carolina Parada byline; three-model introduction; On-Device section
- AIVI-Learning Is Now Powered by Google Gemini RoboticsBoston Dynamics · Primary sourceIntegration introduction; New Asset Support; Zero-Downtime Upgrades; Data Privacy; Looking Ahead