IEEE Robotics & Automation Magazine - December 2017 - 52

1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6

CP
Left-Leg Angle ϕzL in Radians

Right-Leg Angle ϕzR in Radians

Extension

(a)

31.5

Flexion

(b)

Time (s)

(c)

32.5

Normalized Right-Knee Joint
Normalized Right-Hip Joint
33.5

(d)

Figure 3. The different phases during one cycle for the right-knee joint: (a) the start of the extension phase, (b) the extension phase,
(c) the start of the flexion phase, and (d) the flexion phase.

and CP mapping are independent of
the geometry without the need for
IMU
ϑC ϑC
manual pattern tuning. As shown in
afat, R (t )
MR (t )
Table 3, the resulting crank-angle areas
R (t ), L (t )
IMU
Muscle
Moment
Cycling
were different for each geometric setstim, R , stim, L Activation afat, L (t )
Calculation ML (t )
Kinematics
ting, while the CP areas stayed the
IMU
same over all simulations.
The simulation results showed that
IMU
φR (t ), φL(t )
the cadence smoothness (smooth refers
φR (t ), φL(t )
to a low deviation from the mean
cadence value) depends on the seat posiFigure 4. The structure of the simulation model. The pulsewidth and stimulation
tion and inclination. While it was still
o
frequency are used as inputs to generate the crank-angle j C and cadence j C . Attached
possible to induce a cycling motion, sitvirtual inertial sensors generate sensor data based on a mechanics simulation [7].
ting too close to the crank decreased
mechanical model, which was created by using the simula- the cycling quality by means of cadence smoothness because
tion framework SimMechanics. The model covered a of the changed lever-arm relation. Regarding the metabolic
mechanical representation of a cycling motion for the lower efficiency, a tuning of the pattern could be reasonable. With
limbs, providing joint motion, crank angle, cadence, and an optimal stimulation intensity profile over the activation
sensor data of virtual inertial sensors placed on the virtual ranges, the fatigue effects could be delayed, thus yielding lonthighs and shanks. The output of the inertial sensors was ger cycling distances. This is not part of the Cybathlon setused to test the presented methods, i.e., estimating joint ting but will be an objective of future development.
angles and generating stimulation timing.
We carried out a simulation (Figure 5) to validate that the Speed-Dependent Pattern Correction
presented methods can induce a cycling motion without manu- The stimulation pattern described in the previous section
al tuning despite changing geometric parameterizations. assumed that no delay was present between the stimulation
Therefore, we varied the seat position of the model in distance and the resulting joint torque production by the muscles.
and height (with respect to the crank) as well as the inclina- However, torque generation is a dynamic process that must
tion angle of the back of the cyclist. A simple proportional- be considered when activating the muscles during cycling.
integral controller was used to control the mean cycling speed A very rough approximation of the process is a simple time
to 50 rounds per minute (RPM) by controlling the pulsewidth delay of about 130 ms [10]. Therefore, muscles should be
and keeping fstim (t) constant.
stimulated in advance during cycling to guarantee that
Our joint-angle-based approach was able to induce a extension and flexion torques occur exactly at the extension
cycling motion using the proposed pattern despite the geome- and flexion phases of the legs, respectively. This requiretry changes. Furthermore, it was possible to influence the ment could be translated into shifted CP ranges for the
mean cadence to the desired value by changing the stimula- muscle simulation pattern provided in Table 2. A simple
tion intensity. This confirms that the joint-angle estimation linear shift depending on a mean CP time-derivative, as in
52

IEEE ROBOTICS & AUTOMATION MAGAZINE

DECEMBER 2017

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2017