1) The document surveys methods for detecting, isolating, and identifying robot collisions using only proprioceptive sensors.
2) It describes several methods including using estimates of total energy, generalized momentum, joint acceleration, and inverse dynamics. Computational issues with each method are also discussed.
3) Experimental results applying the methods to various robot platforms like a humanoid and Kuka are presented, demonstrating the ability to detect and locate collisions using only internal sensors.
1. Review
Sami Haddadin, Alessandro De Luca, Albu-Schaffer. 2017.
Robot Collisions:
A Survey on Detection, Isolation, and Identification
ModuLabs
강남Dynamics Lab
Hancheol Choi
(babchol@gmail.com)
3. Collision Event Pipeline
• Pre-collision: collision avoidance, minimize impact effects
• Detection: collision TRUE or FALSE
• Isolation: which part being collided
• Identification:
• Classification: accidential, intentional
• Reaction: low-level control
• Post-collision: recognize human’s wish
Independent from context
4. Contribution
• Apply to: A hydraulically driven humanoid, flying robots, KUKA LWR iiwa,
FRANKA EMIKA, ABB YuMi
• Main characteristics of original works
1) Use only proprioceptive sensors
2) Elegant physical motivation: total energy, generalized momentum
3) Rigid, flexible joints
4) Independent from the control method
5) Provide directional and intensity information
• Plus contributions
1) More favorable version of the energy observer that is based on kinetic energy only
2) Elaborate on the momentum-based method, showing sufficient conditions for localizing
the contact force vector using only proprioceptive sensors
3) Compare and rate computational effort, required measurement quantities
4) New experimental results with the link momentum observer for the DLR/KUKA LWR
including its capability for external force estimation as well as collision isolation and
identification.
5) First experimental analysis using the energy observer
6) Simulative analysis of real-world effects on the momentum-based monitoring
• Practical comments
1) Motor/link-side friction – thresholding on the collision detection performance
2) Variant of the momentum observer for flexible robots that does not require any information
on the joint stiffness and uses only motor- and link-side position measurements
6. Rigid Joint Robot
Rigid-body dynamics including external torque
Skew-symmetric property of is equivalent from
Contact jacobian
Total Energy
Generalized momentum
7. Flexible Joint Robot
Joint deflection:
Link dynamics:
Elastic torque(also output of torque sensing device):
Motor dynamics:
Total Energy
Generalized momentum
9. Power Estimation via Energy Observer
Total energy
Derivatives of total energy
Observer dynamics
Intergrate external power estimation
Use kinetic energy instead of total energy
When joint velocity is zero, no collision detected!
Wrenches orthogonal to the contact velocity cannot be detected!
10. Direct Estimation of
It uses joint acceleration. It is double differentiated, so it’s very noisy.
12. Estimation of via Velocity Observer
Actual acceleration
Observer dynamics
Integrate external torque estimation
Nonlinear and coupled because of using inverse of inertia matrix
13. Estimation of via Momentum Observer
Derivatives of momentum
Observer dynamics
Integrate external torque estimation
No need for acceleration
No matter when joint velocity is zero
14. Computational Issues
Recursive Newton-Euler(NE)
ex)
Monified recursive Newton-Euler (NE)
In momentum-observer method, we have to calculate
1) NE and numerical approximation of derivative of the inertia matrix
2) Customized Lagrange dynamics in symbolic form
3) NE and MNE
17. Collision detection of flexible joint robots
Elastic Robots vs. Intrinsically flexible joint robots
Link-side momentum observer
measuring can decouple motor-side friction
Total momentum observer
does not require joint torque sensing
again sensitive to motor-side friction.
Estimating via joint torque sensing or iterative scheme
Second position sensor on the link side, but achieving necessary
resolution for obtaining 𝑞 via numerical differentiation is difficult
Albu-Schaffer, 2007
19. Simulation
1) 𝜃 is quantized with a resolution of 400 increments and filtered with a
first-order filter(cutoff frequency 300Hz)
2) 𝜏𝐽 uniformly distributed noise of ±0.3 Nm, hysteresis Δ𝜏ℎ𝑦𝑠=0.2Nm,
quantization of 12-bit resolution and filtered with a first-order filter(cutoff
frequency 300Hz)
3) Harmonic drive gears produce a ripple torque 𝜏 𝐻𝐷,𝑚𝑎𝑥 = 0.2 Nm at a
frequency, which is twice velocity 𝜃
4) link position and velocity, velocity is filtered with a first-order
filter(cutofffrequency 300Hz)
5) Link-side coulomb friction 𝜏 𝑓 𝑞
= 0.5 Nm
6) Estimated end-effector mass is 50% of the real mass
7) θ, τJ is delayed by td,θ = 1ms, td,τJ
= 1ms
21. Reference
[1] Sami Haddadin, Alessandro De Luca, Albu-Schaffer, “Robot Collisions: A Survey
on Detection, Isolation, and Identification”, TRO, 2017.
[2] A. De Luca and L. Ferrajoli, “A modified Newton-Euler method for dynamic
computations in robot fault detection and control,” in Proc. IEEE
Int. Conf. Robot. Autom., 2009, pp. 3359–3364.
[3] L. Le Tien, A. Albu-Sch¨affer, A. De Luca, and G. Hirzinger, “Friction
observer and compensation for control of robotswith joint torque measurement,”
in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2008, pp. 3789–
3795.
[4] A. Albu-Sch¨affer, C. Ott, and G. Hirzinger, “A unified passivity-based
control framework for position, torque and impedance control of flexible
joint robots,” Int. J. Robot. Res., vol. 26, pp. 23–39, 2007.