Evolutionary Robotics Bipedal Locomotion Question 1: Why only walking and running? Question 2: What (if any) part does the upper body play in bipedal locomotion?

Evolutionary Robotics Bipedal Locomotion Question 1: Why only walking and running? Question 2: What (if any) part does the upper body play in bipedal locomotion? Honda’s ASIMO robot

CS295/CS395/CSYS395 Evolutionary Robotics Bipedal Locomotion Passive Dynamic Locomotion Steve Collins, Martijn Wisse, Andy Ruina, Cornell Hybrid Dynamic Locomotion Martijn Wisse, TU Delft (2004)

Evolutionary Robotics Bipedal Locomotion Research question: Can passive dynamic walking be evolved? How? Could evolved passive dynamic walking be evolved further into hybrid dynamic walking? Vaughan, E., Di Paolo, E., Harvey, I. (2004) The evolution of control and adaptation in a 3D powered passive dynamic walker. In Artificial Life IX, pp. 139-145.

Evolutionary Robotics Bipedal Locomotion Research question: Can passive dynamic walking be evolved? How? Could evolved passive dynamic walking be evolved further into hybrid dynamic walking? Vaughan, E., Di Paolo, E., Harvey, I. (2004) The evolution of control and adaptation in a 3D powered passive dynamic walker. In Artificial Life IX, pp. 139-145.

Vaughan, E., Di Paolo, E., Harvey, I. (2004) The evolution of control and adaptation in a 3D powered passive dynamic walker. In Artificial Life IX, pp. 139-145.

Mw = waist mass Mt = thigh mass Ms = shank mass Mf = foot mass L = leg segment length Xt = thigh mass x-offset Yt = thigh mass y-offset Xs = shank mass x-offset Ys = shank mass y-offset Lf = foot length W = waist radius By = starting hip angle around y-axis Vaughan, E., Di Paolo, E., Harvey, I. (2004) The evolution of control and adaptation in a 3D powered passive dynamic walker. In Artificial Life IX, pp. 139-145. Placing the robot’s body under evolutionary control: evolve not just NN parameters, but body parameters as well.

Ax = ankle spring/damper around x-axis Kx = knee spring/damper Hx = hip spring-/damper Hy = hip spring/damper around y-axis Ay = ankle spring/damper Each spring/damper has two parameters: stiffness and damping coefficient Q: How many body params? Vaughan, E., Di Paolo, E., Harvey, I. (2004) The evolution of control and adaptation in a 3D powered passive dynamic walker. In Artificial Life IX, pp. 139-145. High stiffness Low stiffness Black line: low damping; blue line: high damping Placing the robot’s body under evolutionary control: evolve not just NN parameters, but body parameters as well.

Continuous Time Neural Network (CTNN) Five motors: one attached to each spring/damper. CPG = Two hidden nodes / per motor Sensors: *A = angle sensor what is the angle of the joint? *F = force sensor how far is the spring from its resting length? rot* = rotation about * axis accel*=acceleration about * axis hidden motors sensors

The Fitness Function: Q: Does d, t, x, z, r, or y reward or punish? f: fitness d: distance travelled t: torque used (torque = rotational force) x: hip rotation about x-axis z: hip acceleration about z-axis r: feet rotation about z-axis y: hip rotation about y-axis

Shaping: Evolve passive walking 2. Evolve hybrid walking 3. Evolve sensor-based walking

Shaping: 1. Evolve passive walking Evolve hybrid walking During evolution: Evolve CPGmotor params Evolve body params Add (1/(1+v)) to fitness function v = difference between passive and hybrid velocity Gradually lower declined plane over generations During evaluation: Record time to first heel strike Use this to set CPG frequency 3. Evolve sensor-based walking

Shaping: Evolve passive walking Evolve hybrid walking Evolve sensor-based walking During evolution: reconnect sensors to NN with small connection weights Evolve all NN param weights Evolve body params Use 1/(1+v) During evaluation: Same as before.

Robustness to external perturbations Evolve for another 100 generations. Small random force vectors applied to the robot while it moved (i.e. “wind”). “Developed dynamic mechanisms to adapt to noise.” “When pushed too far to one side, the machine was observed to adjust its foot placement by stepping inward to regain balance.”

Robustness to internal perturbations Evolve further. Introduce “mistakes” when building the body. Random changes are made to the body params

Robustness to CPG perturbations