Human muscles are very unconventional actuators, from an engineering point of view. They are neither pure force generators (like dc electric motors), nor pure motion generators (like stepper motors); rather, they behave like springs, with tunable elastic parameters. This property turns out to be a key element while facing the complexity of motor problems, such as motor redundancy, trajectory formation, negotiation of impacts and motor learning. The built-in compliance of muscles is a winning feature for achieving versatility and robustness, although at the cost of precision. Aimed at reproducing the most salient functional features of natural musc1es, artificial muscle actuators necessarily should target their ideal performances to their biological counterparts. This provides stringent design specifications: linear contractions, robustness, stability, long lifetime, built-in tunable compliance, compactness (low volume-to-power ratio, < 10m /W), high force density (skeletal muscles can generate active stresses from 0.1 to 0.5 MPa), fast response time (mechanical tension of skeletal muscles becomes maximum within 30 ms after neural excitation for a single twitch), high power-to-weight ratio (typical output power densities of human muscles are of the order of 0.1 kW/kg) and high efficiency of energy conversion (biological muscles show values as high as 45-70 %). Some of these performances have been achieved or even exceeded with conventional motor systems, but most of them are still too ambitious to be approached with devices available today.