Redazione RHC : 2 November 2025 09:01
On September 25, Google DeepMind released a video demonstrating how its humanoid platforms handle multi-step, everyday tasks using multimodal reasoning.
In a series of demonstrations, the machines confidently performed sequences of actions, including sorting objects according to predetermined rules.
The intelligence of these systems is based on the Gemini Robotics 1.5 family. Two components work together: the basic model translates visual signals and text messages into specific movements, while the modified Gemini Robotics-ER 1.5 version creates plans and reasoning step-by-step about the current situation, choosing the right sequence of steps.
The so-called banana test clearly demonstrates the progress. Previously, the robot was only required to pick up a banana and place it in a bowl: one command, one result.
Now, the system sorted three different fruits by color and placed them on plates. Jie Tang, a senior researcher at Google DeepMind, demonstrated the experiment; the two-arm system, based on Franka manipulators , completed the entire sequence without a hitch.
Apptronik ‘s Apollo humanoid platform was also tested in a laundry. The machine sorted clothes by shade into two bins: one for whites and one for blacks. After the first successful attempt, the engineers swapped the bins to see if the system would detect the mixup and adjust its actions. Apollo recognized the new arrangement and successfully completed the sorting.
Gemini Robotics 1.5 supports embodied learning: the robot explores its surroundings with its body, sensors, and cameras, and then acts based on its observations . ALOHA 2 was used in most of the scenes, but the same scenarios can also be handled with Apollo and the two-arm Franka system.
Agent-based functions have also been added . For example, the system can be tasked with waste sorting: it will search the internet for local regulations, visually assess each item, assign it to compost, recycling, or garbage, and perform the entire process, from decision to disposal in the appropriate container.
This level of consistency is achieved through the collaboration of two components: one responsible for the path from perception to motion , the other for planning and logic. This architecture makes real-world task execution more intuitive and reliable.
Safety has received particular attention. The robots are trained to proactively assess risks, respect human limitations, and avoid dangerous situations . With the support of specialized teams and the updated ASIMOV testbed , Gemini Robotics-ER 1.5 has achieved a leading position in testing, which should facilitate the accurate implementation of similar systems outside the laboratory.
Redazione