A robot from AI startup Generalist AI has surprised researchers by using a banana as a makeshift brush. The demonstration took place during recent testing after the robot watched a video of sweeping a block into a bowl.
Researchers removed the brush and placed other objects nearby, testing whether the robot could adapt instead of stopping. The robot used a dustpan first and later a banana to move the block into the bowl. The test showed how physical AI can respond when familiar tools suddenly disappear.
Generalist AI trains robots by showing short videos instead of giving separate instructions for every task. The approach aims to help robots transfer learned skills across changing objects and environments. Researchers also tested a two armed robot opening a different purse and removing banknotes.
When the right gripper struggled to hold the money, the robot switched to its left gripper and continued the task. An engineer reacted during the demonstration, saying, “It never did that before.”
Pete Florence, Generalist AI cofounder and CEO, linked the behavior to the early excitement around large language models. “You could take that model and just prompt it to do a new task and it would have a real shot at doing it,” Florence said.
The experiments point toward robots that can improvise instead of following fixed routines. Generalist AI wants robots to build useful physical knowledge from different experiences. The company says its GEN-1 model supports real time robot learning and physical task execution.
Such adaptability could matter in factories, warehouses, homes, and other unpredictable settings. Robots often struggle when objects, tools, or positions differ from training examples.
Generalist AI is trying to narrow that gap through models that learn broader physical patterns. The unusual choice also gives researchers a practical test for general purpose robotics.
Also Read: Alibaba Bets $290M on Next-Gen AI; Aims to Take Over ChatGPT-Style Models