Teaching a robot is the challenge in the career of humanoids, but imagine if it were as simple as explaining or showing what you want, without code, without joystick, without technical commands, that was exactly what the startup Israel
Mentee Robotics has just demonstrated, in a recently released video, the humanoid robot MenteeBot responds to a simple voice command to grab a can of soda.
It listens to the instruction, interprets the meaning, analyzes the environment and executes the task completely autonomously, no remote control, no manual programming, upon receiving the command, in MenteeBot, it first understands the natural language, then its cameras scan the environment, identify the correct object, calculate the route to the table, plan the pressure and adjust the strength of the fingers to pick up the can accurately.
This entire process occurs in a coordinated manner as a single continuous decision. The demonstration itself shows the robot's vision in real time, revealing how perception, planning, navigation and manipulation are integrated into a single intelligent behavior.
The secret is in Mentee's so-called foundation model; instead of training robots with code, the system allows voice learning and demonstration in a similar way to onboarding a new human employee. The robot first learns in simulated environments by trial and error using reinforcement learning, then those skills are transferred to the real world using techniques called Sim2Rea, this reduces risks, accelerates training and allows rapid adaptation to new tasks.
This model combines two types of learning, practical experience focused on the control of actions and data-driven learning that involves language, vision and voice. The result is a robot that learns faster, generalizes better and does not depend on specialists to be taught.
Designed for domestic and industrial tasks, MenteeBot can wash clothes, set the table, organize inventories and transport loads up to 25 kg, operating for hours autonomously with automatic battery exchange.
