Teaching robots with bendy arms to follow instructions

Soft robotic arms bend and squeeze, making them ideal for delicate or confined spaces, but they're harder to control than rigid arms because their deformation is unpredictable. Researchers created ManiSoft, a simulator and benchmark with 6,300 training scenarios and four manipulation tasks designed specifically for soft arms. Current vision-language models perform well on clean tasks but fail badly under randomization, primarily because they can't accurately estimate where a squishy arm actually is or exploit its flexibility to navigate obstacles.