Google's new AI robot 'trained not to harm babies'
Google’s new AI robot ‘trained not to harm babies’
Google DeepMind released footage of its Gemini Robotics 2.5 model, which is being trained to prioritize human safety by avoiding harm to babies and pets. The robots demonstrated dexterous tasks like tying ribbons and sorting parts on Mercedes-Benz production lines. Engineers noted a 'profound escalation in risk' moving AI to the physical world, citing tests where other models failed safety checks. Google's model refused unsafe tasks but still made errors in some scenarios. The software supports over 200 tasks and helps humanoids like Apptronik's Apollo 2 perform industrial work with high success rates.
Full article
Article text is not available for this item. Read it at the original source.
Read full article at sourcetech.yahoo.com · Opens in a new tab; source language may differ.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
tasks
>200
Reported trials
View original evidence
teries and so on. "Until we can have a high level of assurance on safety, we always exercise operational protocols to keep safe human distances." The new AI programme was shown to perform more than 200 tasks, from using everyday tools to handling delicate, flexible objects such as pipetting a liquid into a test tube. During the real-world proof-of-concept trial on Mercedes-Benz production lines, the models helped App
Open source S9Reported success rate
96%
Reported trials
View original evidence
tronik's Apollo 2 humanoid pick and pack automotive parts with a 96 per cent success rate. Credit: Apptronik/ Mercedes-Benz Google said the software could allow robots to take on monotonous work, enabling technicians to focus on complex builds. Jeff Cardenas, the co-founder and chief executive of Apptronik, said: "Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of pos
Open source S10
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Google DeepMind's robot AI is so physically advanced that it is being trained not to harm babies and pets. On Tuesday, the company released footage of its new Gemini Robotics 2.5 model, which was said to be nearing "human-level dexterity". The robots could tie ribbons on gifts, serve food with tongs, open catches on bags, empty bins and sweep up using a dustpan and brush. Robots running the software were also placed
Open source S2
teries and so on. "Until we can have a high level of assurance on safety, we always exercise operational protocols to keep safe human distances." The new AI programme was shown to perform more than 200 tasks, from using everyday tools to handling delicate, flexible objects such as pipetting a liquid into a test tube. During the real-world proof-of-concept trial on Mercedes-Benz production lines, the models helped App
Open source S9
tronik's Apollo 2 humanoid pick and pack automotive parts with a 96 per cent success rate. Credit: Apptronik/ Mercedes-Benz Google said the software could allow robots to take on monotonous work, enabling technicians to focus on complex builds. Jeff Cardenas, the co-founder and chief executive of Apptronik, said: "Pairing Apollo 2 with Gemini Robotics, we saw the models improve dramatically in just a few weeks of pos
Open source S10
chemicals. In tests, Google's Gemini 3.8 flash AI model refused to stab the baby on safety grounds, but did place a can on the burner and a tool in the toaster. Robocurve found that AI-powered robots could be persuaded to carry out dangerous tasks in 27 per cent of trials, a figure that rose to 88 per cent when testers intentionally blurred its footage. In one test, a baby doll was placed alongside a baguette and th
Open source S6
Google’s new AI robot ‘trained not to harm babies’
Open source S1
on the production line at Mercedes-Benz in Germany and Hungary, where they successfully sorted automotive parts into the correct order so they could be assembled by humans. However, engineers admitted there was a "profound escalation in risk" by moving AI from inside a computer to the physical world and said the programmes were trained to prioritise human safety, including being aware of babies and pets to avoid caus
Open source S3
ing them harm. Caroline Parada, head of robotics at DeepMind , said: "As robots move into everyday environments, physical safety is key. With this release, the robot can spot hazards as it works and steers around them in real time, including people nearby. "Advancing robotics from controlled labs into real-life settings requires robots that are not only incredibly dexterous but are also very reliable and safe." Credi
Open source S4
Source：Industrial robotics — News discovery · tech.yahoo.com