Towards Autonomous Robots: Lessons from Biology

Even relatively simple biological organisms are capable of complex behaviors that would be extremely useful for robots. The ability of animals to move through complex terrain, to track prey, to avoid predators, to rapidly extract and act on relevant information, to learn from experience, and to pursue longterm goals, would greatly improve the function of artificial devices. What can we learn from biological organisms? Several important lessons emerge from the design and analysis of biologically-inspired robots. First, biomechanics matters for control. Models of soft-bodied structures demonstrate that biomechanical properties significantly simplify movement and locomotion control. Second, local reflexes can be utilized to simplify higher order control. Incorporating local leg reflexes into a hexapod robot allowed it to successfully negotiate highly irregular and compliant terrain with relatively little direction from higher order control. Third, robust control of behavioral sequencing is important. Studies of cyclic controllers composed of sequences of saddle equilibrium points have suggested designs for robust controllers of locomotion and other rhythmic behaviors that allow arbitrarily long dwell times. These controllers can readily incorporate sensory feedback to generate appropriate behaviors as environmental conditions change over the short term. Fourth, learning paradigms appropriate for specific tasks are crucial for generating appropriate behavior over longer terms. An agent that could sense different forms of food, whose utility was different in different environments, could be evolved to use slower dynamics to rapidly switch its behavior for the appropriate environment based on reinforcement signals. Fifth, truly autonomous robots will need to have internal goals and drives. An insect-like robot given an internal energy detector and an ability to detect and move towards food was capable of sustaining itself for an extended period of time in an artificial environment. More challenging goals may be pursued if agents have more complex internal states and the ability to learn from experience. Finally, future agents may benefit from a deeper understanding of the key biological processes that work together to generate organisms adapted to complex environments - evolution, development, and plasticity - and the importance of starting from good initial conditions.