GPC: Steppy

Climbing an Unknown-Height Step Using Proprioception and Compliance

Step Climbingogv|mp4|webm

Climbing a 2cm step (actual speed)

While a number of stair climbing humanoids have already been developed, in most cases the robot was given the step parameters (height, depth, etc) a-priori. Here we only allow the robot to assume that the step height is somewhere in the range 0 to 35mm, about 10% of the robot height. Further, the only sensing the robot may use is positional proprioception at its joints.

An experiment with about 100 trials yielded a 90% success rate, with success defined as termination of motion with the robot in stable double support on the upper platform. The robot autonomously assembles pre-defined motion plan fragments on-line based only on position feedback from its joint actuators. The impedance of the actuators is selectively lowered by decreasing position loop proportional gain to allow collision to backdrive the joints and thus provide information about the colliding geometry.

The hardware is based the Robotis Bioloid (~36cm tall, ~2kg). We added some components to stiffen the hip yaw DoF and coated the soles (but not the edges) of the feet with several layers of silicone conformal coating for stiction on the climbing platforms. We also added external fans which reduce the maximum temperature of the servos by about 30°C.

We implemented our own firmware which runs on the main controller of the Bioloid and exchanges afferent sensor data and efferent commands for all DoF to/from a remote-brain workstation at 10Hz. The workstation can run an interactive GUI for manipulating the robot and editing motion sequences in addition to autonomously executing adaptive step sequences based on proprioceptive feedback as described above.

Modeling

What limits the amount of uncertainty—here, variation in step height—that can be tolerated with this approach? We explored this question by using virtual articulations (virtual kinematic links and joints) to model the foot contacts, the hip compliance, and the uncertain step height. This effectively gives a low-dimensional representation of salient aspects of the full physical system. Though it is only an approximation, we found that predictions of the maximum tolerable step height compared well with physical experiments. For further details see Chapter 6 and Appendix L of Prof. Vona’s Dissertation.
Virtual Articulation Modelogv|mp4|webm

Quasi-static simulation where the step height is modeled as a virtual prismatic joint and the contacts of each foot are also modeled with virtual joints. The minimum detectable step is determined by the first contact of the right heel, which occurs when the associated virtual joint hits a limit. The maximum detectable step is limited by the tip angle of the robot, detectable by the rotation of the left foot virtual contact joint. Simulation results for these limits compare well with actual hardware experiments, even though the model is greatly simplified.