The HUBO FX-1 is a two-legged robot with a cockpit area to carry passengers. Besides being very entertaining to ride around in, the robot could also be used to help wheelchair-bound people regain mobility and overcome tough obstacles such as staircases.
Research for this robots has so far focussed on generating a walking pattern while controlling balance and reducing vibrations with parameters optimized for a passenger of a certain fixed weight. However, if the passenger is much lighter or much heavier than the predefined weight, the walking performance and stability go down.
To overcome this problem, Kim et al. propose a new approach that adapts to the weight of the passenger measured by using the force/torque sensors at the feet of the robot. Results show that their approach achieves better walking performance than the original non-adaptive approach.
If you’ve never seen a video of springboks gracefully pronking, have a look below.
Pronking is a gait where all legs are used in synchrony, usually resulting in relatively slow speeds but long flight phases and large jumping heights. Such jumps might be interesting for robots to move around in cluttered environments. The risk is that the robot falls forward or backward if not controlled correctly (pitch control).
For this purpose, Ankaralı et al. propose a special type of feedback controller that has two levels. The top-level takes as an input the desired speed and jump height of the robot. This information is given to a “template” of the robot motion based on the “Spring-Loaded Inverted Pendulum”. A low-level controller then attempts to force the dynamics of the robot to mimic the template as closely as possible.
Experiments were done in simulation on a realistic model of the RHex six-legged robot (see video above). Results show that the user can easily control jump height and forward speed and that the gait is robust to sensor and actuator noise.
The typical way to make a bipedal robot walk is to actuate its leg joints, strap a bunch of sensors to measure its state and add a tight control loop to make sure it is performing the desired steps.
In a radically different approach, passive dynamic walkers can step down slopes without the need for sensing, control or energy. Their driving force comes from gravitation pushing them down the hill. If well designed, and started with adequate initial conditions, the walker will reach a rhythmic and stable walking gait that prevents it from falling on its nose.
Of course, always walking downhill is hardly a viable solution. To make robots walk on level ground, Dong et al. propose to trick the robot into thinking it’s walking on a slope. This is done by extending the back leg of the robot (stance leg) while shortening its front leg (swing leg) before it hits the ground as shown in the figure below (steps I through IV).
The authors propose an analytical model to predict the energy efficiency and speed of the walker based on easy to tune parameters. The result is an energy efficient walker that can move at high speeds. To validate their model, experiments were done on the real walker below. The robot was able to top at a full 1.12 m/s speed, or 4.48leg/s, which is the fastest walking gate demonstrated so far. The leg length was changed by bending and unbending the knee joints.
Animal walking is thought to be driven by rhythmic signals sent through the spinal cord. These signals are translated to motions of the limbs. For a bipedal walker, such patterns would force leg swings and foot contacts to be alternated so as to achieve stable walking. By using similar mechanisms, roboticists hope to generate walking gates that do not require any complex modeling or computation.
Along these lines, Aoi et al. consider stable walking with a five-link biped robot. The links represent the femur and tibia of both legs and torso as shown in the video below. The robot is driven by a Central Pattern Generator (CPG) that uses one oscillator to generate the rhythmic signals. As a first step, they investigate what parameters lead to stable walking when no sensory feedback is used (open-loop). Important parameters include walking speed, knee amplitude, and distribution of mass. In a second step, the robot is able to detect when its foot hits the ground and use that information to reset the oscillator. By reacting to its environment, the robot is therefor able to adapt its walking and achieve better stability. Finally, controller parameters for the walker are optimized to fully exploit the interactions between robot dynamics, oscillator dynamics and the environment.
Imagine walking on a flat surface with your eyes blinded. If the slope below your feet changes, you’ll most likely change your posture to keep moving. To explain this, an idea from the 1950s says that we can predict the sensation that will be produced by a motor command sent by our central nervous system. We can therefore tell apart sensations that are due to our own motion and sensations due to external stimuli. When the expected sensation doesn’t match the sensory input, we change our behavior to compensate.
In work by Schröder-Schetelig et al., a robotic walker uses this idea to stay on its two feet. More precisely, the robot uses a neural network (which is a type of controller) to send commands to hip-joint and knee-joint motors such that the robot is able to walk on flat terrain. These motor commands are then copied (efference copy) and fed to a second neural network that captures the internal model of the robot. This model predicts the acceleration the robot should feel given its motor command and current state. If the acceleration is larger than expected, the robot is probably going downhill and should lean back to slow down. Likewise, if the acceleration is lower, the robot is going uphill and should lean forward. Leaning backward and forward is performed by moving a mass that represents the upper body of the robot and is controlled by a third neural network that takes as an input the robot’s predicted acceleration and the measured acceleration given by an accelerometer.
Experiments shown in the video below were conducted on Runbot, a 23cm bipedal robot that is physically constrained to a circular path of 1m radius and can not perform sideway movements. Results show the robot successfully climbing a changing slope.
In the future, Schröder-Schetelig et al. hope to refine the internal model of Runbot, make it climb even steeper slopes and adapt to new and unforeseen environments.
