Literature
Display
← All papers

Robot Control based on Motor Primitives — A Comparison of Two Approaches

Nah, Moses C. · Lachner, Johannes · Hogan, Neville

· The International Journal of Robotics Research ↗

Robotics9 min text · CondensedOriginal paper ↗
Contents

At a glance

Dynamic movement primitives (DMPs) parameterize motion trajectories; elementary dynamic actions (EDAs) combine submovements, oscillations, and mechanical impedances. This edition follows the requested 2023 preprint, which compares them through eight simulated control examples. The later journal publication is linked above. The comparison emphasizes trajectory reproduction, contact behavior, singularities, and composition rather than ranking one method universally better.

1 Introduction

Both approaches build complex movement from reusable primitives, but at different levels. DMPs learn a forcing function within an attractor-based trajectory generator. EDAs separate a virtual movement from the mechanical response connecting it to the robot. The paper compares particular torque-control implementations and suggests combining their strengths; its findings should not be read as restrictions on every possible DMP controller.

2 Theory

The theory begins with one degree of freedom (DOF) before extending the constructions to robot joints and end-effector motion.

2.1 Dynamic Movement Primitives

A DMP consists of a phase-generating canonical system, a learned nonlinear forcing term, and a transformation system producing motion. Discrete and rhythmic variants use different phase and basis functions.

2.1.1 Canonical System

Discrete phase decays exponentially toward zero; rhythmic phase advances around a cycle. A time constant controls movement speed and duration.

Equation 1
τ​s˙​(t)={−αs​s​(t) Discrete 1 Rhythmic\displaystyle\tau\dot{s}(t)=\begin{cases}-\alpha_{s}s(t)&\text{ \footnotesize Discrete }\\ 1&\text{ \footnotesize Rhythmic}\end{cases} (1)

2.1.2 Nonlinear Forcing Term

Weighted Gaussian bases shape discrete movements, while periodic von Mises bases shape rhythmic ones. Start/goal displacement or oscillation amplitude scales the forcing term.

Equations 2–3
f⁡(s⁡(t))={∑i=1Nwi​ϕi​(s⁡(t))∑i=1Nϕi​(s⁡(t))​s​(t)​(g−y0) Discrete ∑i=1Nwi​ϕi​(s⁡(t))∑i=1Nϕi​(s⁡(t))​r Rhythmic\displaystyle f(s(t))=\begin{cases}\frac{\sum_{i=1}^{N}w_{i}\phi_{i}(s(t))}{\sum_{i=1}^{N}\phi_{i}(s(t))}s(t)(g-y_{0})&\text{ \footnotesize Discrete }\\ \frac{\sum_{i=1}^{N}w_{i}\phi_{i}(s(t))}{\sum_{i=1}^{N}\phi_{i}(s(t))}r&\text{ \footnotesize Rhythmic }\end{cases} (2)
ϕi​(s⁡(t))={exp⁡{−hi​(s⁡(t)−ci)2}Discrete exp⁡{hi​(cos⁡(s⁡(t)−ci)−1)}Rhythmic\displaystyle\phi_{i}(s(t))=\begin{cases}\exp\big\{-h_{i}(s(t)-c_{i})^{2}\big\}&\text{\footnotesize Discrete }\\ \exp\big\{h_{i}(\cos(s(t)-c_{i})-1)\big\}&\text{\footnotesize Rhythmic }\end{cases} (3)

2.1.3 Transformation System

A second-order transformation system resembles a damped spring driven by the learned force. With a vanishing discrete forcing term and suitable positive gains, it converges toward the goal. Phase supplies a reusable progression variable rather than an explicit time-indexed trajectory.

Equation 4
τ​y˙​(t)\displaystyle\tau\dot{y}(t) =z⁡(t)\displaystyle=z(t) (4)
τ​z˙​(t)\displaystyle\tau\dot{z}(t) =αz​{βz​(g−y⁡(t))−z⁡(t)}+f⁡(s⁡(t))\displaystyle=\alpha_{z}\{\beta_{z}(g-y(t))-z(t)\}+f(s(t))

2.1.4 Imitation Learning

Demonstrated positions, velocities, and accelerations determine a target forcing function. Locally weighted or least-squares regression fits its basis weights. The resulting motion can be rescaled in time and space, and multiple joints share one phase variable. Reproduction accuracy depends on the representation, data, and subsequent robot controller; fitting a forcing function does not itself guarantee perfect physical tracking.

Equations 5
ft​a​r​g​e​t​(s⁡(t))=τ2​y¨d​e​s​(t)+αz​τ​y˙d​e​s​(t)+αz​βz​(yd​e​s​(t)−g)f_{target}(s(t))=\tau^{2}\ddot{y}_{des}(t)+\alpha_{z}\tau\dot{y}_{des}(t)+\alpha_{z}\beta_{z}(y_{des}(t)-g)
wi∗=𝐚T​Φi​𝐟t​a​r​g​e​t𝐚T​Φi​𝐚w_{i}^{*}=\frac{\mathbf{a}^{\text{T}}\Phi_{i}\mathbf{f}_{target}}{\mathbf{a}^{\text{T}}\Phi_{i}\mathbf{a}} (5)
𝐚\displaystyle\mathbf{a} =[a1a2aP]𝐟t​a​r​g​e​t=[ft​a​r​g​e​t,1ft​a​r​g​e​t,2ft​a​r​g​e​t,P]\displaystyle=\begin{bmatrix}a_{1}\\ a_{2}\\ \vdots\\ a_{P}\\ \end{bmatrix}\quad\mathbf{f}_{target}=\begin{bmatrix}f_{target,1}\\ f_{target,2}\\ \vdots\\ f_{target,P}\\ \end{bmatrix}
Φi\displaystyle\Phi_{i} =[ϕi​(s⁡(t1))0ϕi​(s⁡(t2))⋱0ϕi​(s⁡(tP))]\displaystyle=\begin{bmatrix}\phi_{i}(s(t_{1}))&&&0\\ &\phi_{i}(s(t_{2}))&&\\ &&\ddots&\\ 0&&&\phi_{i}(s(t_{P}))\\ \end{bmatrix}
ai≡a⁡(ti)\displaystyle a_{i}\equiv a(t_{i}) ={s⁡(ti)​(g−y0) Discrete r Rhythmic\displaystyle=\begin{cases}s(t_{i})(g-y_{0})&\text{ \footnotesize Discrete }\\ r&\text{ \footnotesize Rhythmic }\end{cases}
ft​a​r​g​e​t,i\displaystyle f_{target,i} =τ2​y¨d​e​s​(ti)+αz​τ​y˙d​e​s​(ti)+αz​βz​(yd​e​s​(ti)−g)\displaystyle=\tau^{2}\ddot{y}_{des}(t_{i})+\alpha_{z}\tau\dot{y}_{des}(t_{i})+\alpha_{z}\beta_{z}(y_{des}(t_{i})-g)

2.2 Elementary Dynamic Actions

EDAs use three primitive types: finite submovements, periodic oscillations, and mechanical impedances that govern interaction.

Original figure
Fig. 1: The three primitives of Elementary Dynamic Actions (EDAs). Submovements and oscillations correspond to kinematic primitives and mechanical impedances manage physical interaction.

2.2.1 Submovements

A submovement is a smooth finite-duration displacement with a single-peaked velocity profile. Its amplitude, duration, timing, and shape parameterize a discrete motion.

Equation
x˙0​(t)=v​σ^​(t)\dot{x}_{0}(t)=v\ \hat{\sigma}(t)

2.2.2 Oscillations

An oscillation supplies a smooth periodic virtual trajectory. More general almost-periodic motions are possible, but the examples use ordinary periodic functions.

Equation
∀t>0:∃T>0:x0​(t)=x0​(t+T)\forall t>0:~~\exists T>0:~~x_{0}(t)=x_{0}(t+T)

2.2.3 Mechanical Impedances

Mechanical impedance maps displacement from a virtual trajectory to interaction force, including stiffness, damping, and more general dynamics. Compatible impedance operators can be superimposed, with Jacobian-transpose maps converting task-space forces to joint torques. Stability and passivity depend on the operators and interconnection assumptions.

Equations 6–7
Z:Δ​x​(t)⟶F⁡(t)Z:\Delta x(t)\longrightarrow F(t) (6)
Z=∑ZiZ=\sum Z_{i} (7)

2.2.4 Norton Equivalent Network Model

The Norton-equivalent view separates the virtual trajectory from the impedance connecting it to the actual robot motion. Neither position nor force is directly prescribed: both arise from interaction with the environment. Submovements and oscillations can be added at the trajectory level, while impedance components can be combined at the force level.

Original figure
Fig. 2: The three Elementary Dynamic Actions (EDAs) combined using a Norton equivalent network model. The virtual trajectory x0​(t)x_{0}(t) (yellow box) consists of submovements (orange box) and/or oscillations (blue box), and mechanical impedance ZZ (green box) regulates the interactive dynamics.

3 Comparison of the Two Approaches

Eight MuJoCo examples compare reaching, contact, avoidance, rhythmic motion, combined motions, sequencing, and redundancy. The models assume ideal torque actuation and gravity compensation; end-effector orientation is omitted. These are explanatory simulations, not a hardware benchmark or statistical evaluation.

Equation 8
𝐌⁡(𝐪)​𝐪¨+𝐂⁡(𝐪,𝐪˙)​𝐪˙+𝐆⁡(𝐪)=𝝉i​n​(t)+𝝉e​x​t​(t)\mathbf{M}(\mathbf{q})\ddot{\mathbf{q}}+\mathbf{C}(\mathbf{q},\dot{\mathbf{q}})\dot{\mathbf{q}}+\mathbf{G}(\mathbf{q})=\bm{\tau}_{in}(t)+\bm{\tau}_{ext}(t) (8)

3.1 The Existence of an Inverse Dynamics Model

The implementations differ in how a generated motion becomes a torque command.

3.1.1 Dynamic Movement Primitives

The chosen DMP implementation converts desired kinematics to feedforward torque through inverse dynamics, adding feedback for disturbance rejection. Task-space trajectories additionally require inverse kinematics or related feedback mappings. Exact tracking claims assume an accurate model and favorable operating conditions.

3.1.2 Elementary Dynamic Actions

The EDA implementation maps measured displacement and velocity through joint/task impedances to torque. It needs state measurements and forward kinematics but not a full inverse dynamics model for these examples. More advanced impedance shaping may require additional model knowledge.

Equation
𝝉i​n​(t)=∑i𝐉​(𝐪⁡(t))T​𝐙p,i​(Δ​𝐩​(t),t)+∑i𝐙q,i​(Δ​𝐪​(t),t)\bm{\tau}_{in}(t)=\sum_{i}\mathbf{J}(\mathbf{q}(t))^{\text{T}}\mathbf{Z}_{p,i}(\Delta\mathbf{p}(t),t)+\sum_{i}\mathbf{Z}_{q,i}(\Delta\mathbf{q}(t),t)

3.2 A Goal-directed Discrete Movement in Joint-space

The first example reaches a target joint configuration along a specified motion.

3.2.1 Dynamic Movement Primitives

Each joint has a DMP transformation system synchronized by a common phase. A learned force reproduces the desired trajectory, whose position, velocity, and acceleration feed the inverse dynamics controller.

Equation 9
τ2​𝐪¨​(t)=−αz​βz​{𝐪⁡(t)−𝐠}−αz​τ​𝐪˙​(t)+𝐟⁡(s⁡(t))τ​s˙​(t)=−αs​s​(t)\displaystyle\begin{aligned} \tau^{2}\ddot{\mathbf{q}}(t)&=-\alpha_{z}\beta_{z}\{\mathbf{q}(t)-\mathbf{g}\}-\alpha_{z}\tau\dot{\mathbf{q}}(t)+\mathbf{f}(s(t))\\ \tau\dot{s}(t)&=-\alpha_{s}s(t)\end{aligned} (9)

3.2.2 Elementary Dynamic Actions

Joint-space stiffness and damping connect the robot to a finite virtual submovement ending at the goal. Under the stated conditions, the configuration converges to that goal even when it does not track the virtual trajectory exactly.

Equation 10
𝝉i​n​(t)=𝐊q​{𝐪0​(t)−𝐪}+𝐁q​{𝐪˙0​(t)−𝐪˙}\bm{\tau}_{in}(t)=\mathbf{K}_{q}\{\mathbf{q}_{0}(t)-\mathbf{q}\}+\mathbf{B}_{q}\{\dot{\mathbf{q}}_{0}(t)-\dot{\mathbf{q}}\} (10)

3.2.3 Simulation Example

A two-link planar arm follows minimum-jerk joint movements. The ideal-model DMP controller closely tracks the specified motion; the impedance controller exhibits transient tracking error but reaches the target without inverse dynamics.

Equation 11
𝐪d​e​s​(t),𝐪0​(t)={𝐪i+(𝐠−𝐪i)​fM​J​T​(t) 0≤t<T𝐠 T≤tfM​J​T​(t)=10​(tT)3−15​(tT)4+6​(tT)5\displaystyle\begin{split}&\mathbf{q}_{des}(t),\mathbf{q}_{0}(t)=\begin{cases}\mathbf{q}_{i}+(\mathbf{g}-\mathbf{q}_{i})f_{MJT}(t)&\text{ $0\leq t<T$}\\ \mathbf{g}&\text{ $T\leq t$}\end{cases}\\ &\quad\quad\quad f_{MJT}(t)=10\Big(\frac{t}{T}\Big)^{3}-15\Big(\frac{t}{T}\Big)^{4}+6\Big(\frac{t}{T}\Big)^{5}\end{split} (11)
Original figure
Fig. 3: A goal-directed discrete movement in joint-space for Dynamic Movement Primitives (DMPs, blue) and Elementary Dynamic Actions (EDAs, orange) (Section 3.2.3). The black dotted line is a minimum-jerk trajectory (Equation (11)). Goal location: 𝐠=[1,1]\mathbf{g}=[1,1] rad. Parameters of minimum-jerk trajectory: 𝐪i=[0,0]\mathbf{q}_{i}=[0,0] rad, T=1T=1s. Parameters of DMPs: αz=10\alpha_{z}=10, βz=2.5\beta_{z}=2.5, αs=1.0\alpha_{s}=1.0, τ=T\tau=T, N=50N=50, P=100P=100, ci=exp(−αs(i−1)/(N−1))c_{i}=\exp(-\alpha_{s}(i-1)/(N-1)) for i∈[1,2,⋯,N]i\in[1,2,\cdots,N], hi=1/(ci+1−ci)2h_{i}=1/(c_{i+1}-c_{i})^{2} for i∈[1,2,⋯,N−1]i\in[1,2,\cdots,N-1], hN=hN−1h_{N}=h_{N-1}. Parameters of EDAs: 𝐊q=150​𝐈2\mathbf{K}_{q}=150\mathbf{I}_{2} N⋅\cdotm/rad, 𝐁q=50​𝐈2\mathbf{B}_{q}=50\mathbf{I}_{2} N⋅\cdotm⋅\cdots/rad, where 𝐈2∈ℝ2×2\mathbf{I}_{2}\in\mathbb{R}^{2\times 2} is an identity matrix. For DMP, perfect tracking was achieved. For EDA, a non-negligible tracking error was observed.

3.3 A Goal-directed Discrete Movement in Task-space

The next example reaches in Cartesian task space, including a target at a fully extended singular configuration.

Refer to caption
Fig. 4: A goal-directed discrete movement in task-space for (A, C, E) Dynamic Movement Primitives (DMPs, blue) and (B, D, E) Elementary Dynamic Actions (EDAs, orange) (Section 3.3.3). (A, B, C, D) Goal-directed discrete movements in a direction along the positive YY-axis direction, but (C, D) reached kinematic singularity. The black dotted line is a minimum-jerk trajectory (Equation (14)). (E) Time tt vs. σm​i​n​((𝐉⁡(𝐪)))/σm​a​x​((𝐉⁡(𝐪)))\sigma_{min}((\mathbf{J(q)}))/\sigma_{max}((\mathbf{J(q)})) of both robots for (C, D), where σ\sigma is a singular value of the matrix. For DMPs, the MuJoCo simulation halted near 0.99s. Parameters of a minimum-jerk trajectory: (A, B) 𝐩i=[0.0,0.52]\mathbf{p}_{i}=[0.0,0.52]m, 𝐠=[0.0,1.72]\mathbf{g}=[0.0,1.72]m, T=1.0T=1.0s (C, D) 𝐩i=[0.0,0.52]\mathbf{p}_{i}=[0.0,0.52]m, 𝐠=[0.0,2.00]\mathbf{g}=[0.0,2.00]m, T=1.0T=1.0s, where 𝐩i\mathbf{p}_{i} is the initial end-effector position. Parameters of DMPs are identical to those in Figure 3. Parameters of EDAs: 𝐊p=60​𝐈2\mathbf{K}_{p}=60\mathbf{I}_{2} N/m, 𝐁p=20​𝐈2\mathbf{B}_{p}=20\mathbf{I}_{2} N⋅\cdots/m. For DMP, perfect tracking was achieved when the motion was not near a kinematic singularity. However for (C) when the goal 𝐠\mathbf{g} was at a kinematic singularty, the approach failed to achieve the task. For EDA, a non-negligible tracking error existed, but goal-directed discrete movements were achieved for both (A) and (C). The approach based on EDA was not only stable near and at a kinematic singularity but even passed through it (E, near t=1.0t=1.0 and t=2.4t=2.4, depicted as grey circle markers).

.

3.3.1 Dynamic Movement Primitives

DMP-generated Cartesian kinematics are mapped to joints using the Jacobian inverse before inverse dynamics is applied. The simple inverse requires a square, nonsingular Jacobian; redundancy and singularities need additional handling.

Equations 12
τ2​𝐩¨​(t)=−αz​βz​(𝐩⁡(t)−𝐠)−αz​τ​𝐩˙​(t)+𝐟⁡(s⁡(t))τ​s˙​(t)=−αs​s​(t)\displaystyle\begin{aligned} \tau^{2}\ddot{\mathbf{p}}(t)&=-\alpha_{z}\beta_{z}(\mathbf{p}(t)-\mathbf{g})-\alpha_{z}\tau\dot{\mathbf{p}}(t)+\mathbf{f}(s(t))\\ \tau\dot{s}(t)&=-\alpha_{s}s(t)\end{aligned} (12)
𝐪d​e​s\displaystyle\mathbf{q}_{des} =𝐡−1​(𝐩d​e​s)\displaystyle=\mathbf{h}^{-1}(\mathbf{p}_{des})
𝐪˙d​e​s\displaystyle\dot{\mathbf{q}}_{des} =𝐉​(𝐪d​e​s)−1​𝐩˙d​e​s\displaystyle=\mathbf{J}(\mathbf{q}_{des})^{-1}\dot{\mathbf{p}}_{des}
𝐪¨d​e​s\displaystyle\ddot{\mathbf{q}}_{des} =𝐉​(𝐪d​e​s)−1​{𝐩¨d​e​s−𝐉˙​(𝐪d​e​s)​𝐪˙d​e​s}\displaystyle=\mathbf{J}(\mathbf{q}_{des})^{-1}\{\ddot{\mathbf{p}}_{des}-\dot{\mathbf{J}}(\mathbf{q}_{des})\dot{\mathbf{q}}_{des}\}

3.3.2 Elementary Dynamic Actions

Task-space stiffness and damping produce a Cartesian force mapped to torque by the Jacobian transpose. This avoids Jacobian inversion, so the command remains defined at a singularity. Goal convergence still relies on the controller’s stated assumptions.

Equation 13
𝝉i​n​(t)=𝐉​(𝐪)T​[𝐊p​{𝐩0​(t)−𝐩}+𝐁p​{𝐩˙0​(t)−𝐩˙}]\bm{\tau}_{in}(t)=\mathbf{J}(\mathbf{q})^{\text{T}}\big[\mathbf{K}_{p}\{\mathbf{p}_{0}(t)-\mathbf{p}\}+\mathbf{B}_{p}\{\dot{\mathbf{p}}_{0}(t)-\dot{\mathbf{p}}\}\big] (13)

3.3.3 Simulation Example

Both controllers reach ordinary targets, but the unregularized DMP inverse becomes numerically unstable near full extension. The EDA controller crosses the singularity without numerical failure, though it oscillates because task-space damping cannot dissipate all joint motion there. Joint damping can reduce this oscillation; damped inverse methods and feedback can address the DMP implementation’s difficulty.

Equation 14
𝐩d​e​s​(t),𝐩0​(t)={𝐩i+(𝐠−𝐩i)​fM​J​T​(t) 0≤t<T𝐠 T≤tfM​J​T​(t)=10​(tT)3−15​(tT)4+6​(tT)5\displaystyle\begin{split}&\mathbf{p}_{des}(t),\mathbf{p}_{0}(t)=\begin{cases}\mathbf{p}_{i}+(\mathbf{g}-\mathbf{p}_{i})f_{MJT}(t)&\text{ $0\leq t<T$}\\ \mathbf{g}&\text{ $T\leq t$}\end{cases}\\ &\quad\quad\quad f_{MJT}(t)=10\Big(\frac{t}{T}\Big)^{3}-15\Big(\frac{t}{T}\Big)^{4}+6\Big(\frac{t}{T}\Big)^{5}\end{split} (14)
Refer to caption
Fig. 5: A goal-directed discrete movement with unexpected physical contact for (A, B, C, D) Dynamic Movement Primitives (DMPs, blue) and (E, F, G, H) Elementary Dynamic Actions (EDAs, orange) (Section 3.4.3). For DMP: (A) Moment of contact. (A→\rightarrowB) DMP without PD controller (A→\rightarrowC) DMP with PD controller. (D) Time vs. YY-coordinate of the end-effector for A→\rightarrowB (dotted line) and A→\rightarrowC (filled line). For EDA: (E) Moment of contact. (E→\rightarrowF) EDA without impedance modulation (E→\rightarrowG) EDA with impedance modulation. (H) Time vs. YY-coordinate of the end-effector for E→\rightarrowF (dotted line) and E→\rightarrowG (filled line). With impedance modulation of E→\rightarrowG, a slower movement than E→\rightarrowF was achieved. (D, H) The black dotted lines are a minimum-jerk trajectory (Equation (14)). Parameters of the minimum-jerk trajectory, DMPs and EDAs are identical to those in Figure 4. Parameters of the PD controller of DMPs (Equation (15)): 𝐊q=50​𝐈2\mathbf{K}_{q}=50\mathbf{I}_{2} N⋅\cdotm/rad, 𝐁q=30​𝐈2\mathbf{B}_{q}=30\mathbf{I}_{2} N⋅\cdotm⋅\cdots/rad. Smaller gain values than Figure 3 were chosen to achieve a relatively low-gain PD controller as recommended by (Schaal et al., 2007; Pastor et al., 2013). Parameters of EDAs: ℒm​a​x=2.5\mathcal{L}_{max}=2.5J.

3.4 Tasks with Unexpected Physical Contact

An obstacle unexpectedly interrupts a reaching movement and is later removed.

3.4.1 Dynamic Movement Primitives

A low-gain proportional–derivative (PD) feedback torque is added to the DMP feedforward command. This helps recover after contact while reusing the learned trajectory. It is equivalent to one simple joint impedance only under ideal torque actuation.

Equation 15
𝝉f​b​(t)=𝐊q​{𝐪d​e​s​(t)−𝐪}+𝐁q​{𝐪˙d​e​s​(t)−𝐪˙}𝝉i​n​(t)=𝝉f​f​(t)+𝝉f​b​(t)\displaystyle\begin{split}\bm{\tau}_{fb}(t)&=\mathbf{K}_{q}\{\mathbf{q}_{des}(t)-\mathbf{q}\}+\mathbf{B}_{q}\{\dot{\mathbf{q}}_{des}(t)-\dot{\mathbf{q}}\}\\ \bm{\tau}_{in}(t)&=\bm{\tau}_{ff}(t)+\bm{\tau}_{fb}(t)\end{split} (15)

3.4.2 Elementary Dynamic Actions

The EDA controller changes stiffness and damping through an energy-dependent scaling factor. The example bounds the combined kinetic and elastic energy under its specified conditions; if kinetic energy already exceeds the threshold, scaling is clipped to prevent negative stiffness. This simplified demonstration is not a general contact-safety certification.

Equations 16
𝝉i​n=𝐉​(𝐪)T​[𝐊p′​(λ)​{𝐩0−𝐩}+𝐁p′​(λ)​{𝐩˙0−𝐩˙}]\bm{\tau}_{in}=\mathbf{J}(\mathbf{q})^{\text{T}}\big[\mathbf{K}_{p}^{\prime}(\lambda)\{\mathbf{p}_{0}-\mathbf{p}\}+\mathbf{B}_{p}^{\prime}(\lambda)\{\dot{\mathbf{p}}_{0}-\dot{\mathbf{p}}\}\big] (16)
λ={1if​ℒc​(t,λ)≤ℒmaxmax⁡(1𝒰​(ℒmax−𝒯),0)if​ℒc​(t,λ)>ℒmax\lambda=\begin{cases}\hfil 1&\text{if}\ \mathcal{L}_{c}(t,\lambda)\leq\mathcal{L}_{\textrm{max}}\\ \hfil\max\bigg(\frac{1}{\mathcal{U}}(\mathcal{L}_{\textrm{max}}-\mathcal{T}),~~0\bigg)&\text{if}\ \mathcal{L}_{c}(t,\lambda)>\mathcal{L}_{\textrm{max}}\end{cases}
𝒯⁡(𝐪,𝐪˙)\displaystyle\mathcal{T}(\mathbf{q},\dot{\mathbf{q}}) =12​𝐪˙T​𝐌​(𝐪)​𝐪˙\displaystyle=\frac{1}{2}\dot{\mathbf{q}}^{\text{T}}\mathbf{M}(\mathbf{q})\dot{\mathbf{q}}
𝒰⁡(Δ​𝐩,𝐊p)\displaystyle\mathcal{U}(\Delta\mathbf{p},\mathbf{K}_{p}) =12​Δ​𝐩T​𝐊p​Δ​𝐩\displaystyle=\frac{1}{2}\Delta\mathbf{p}^{\text{T}}\mathbf{K}_{p}\Delta\mathbf{p}
ℒc​(t,λ)\displaystyle\mathcal{L}_{c}(t,\lambda) =𝒯⁡(𝐪,𝐪˙)+λ​𝒰​(Δ​𝐩)\displaystyle=\mathcal{T}(\mathbf{q},\dot{\mathbf{q}})+\lambda\mathcal{U}(\Delta\mathbf{p})

3.4.3 Simulation Example

The feedforward-only DMP controller rebounds and misses the goal; adding feedback permits recovery. Both EDA variants eventually reach the goal, but energy modulation reduces the large acceleration released when the obstacle is removed.

Original figure
Fig. 6: Simulation results using Elementary Dynamic Actions (EDAs) with impedance modulation via energy regulation (Equation (16)) (Figure 5) (Section 3.4.3). (Top) Time tt vs. λ\lambda of the controller. After the first contact, a sudden change in λ\lambda occurred. (Bottom) Time vs. ℒc\mathcal{L}_{c} of the controller. Since the zero-force trajectory continued during contact, the decrease of λ\lambda limited ℒc\mathcal{L}_{c} to a maximal value of 2.52.5J. As expected, ℒc\mathcal{L}_{c} did not exceed ℒm​a​x=2.5\mathcal{L}_{max}=2.5J.

3.5 Obstacle Avoidance

A known stationary obstacle lies along the nominal reaching path.

3.5.1 Dynamic Movement Primitives

A coupling term steers the DMP acceleration away from the obstacle while retaining the learned forcing weights. This illustrates online trajectory modulation rather than refitting a demonstration.

Equations 17
τ2​𝐩¨​(t)\displaystyle\tau^{2}\ddot{\mathbf{p}}(t) =−αz​βz​{𝐩⁡(t)−𝐠}−αz​τ​𝐩˙​(t)+𝐟⁡(s⁡(t))+𝐂t​(t)\displaystyle=-\alpha_{z}\beta_{z}\{\mathbf{p}(t)-\mathbf{g}\}-\alpha_{z}\tau\dot{\mathbf{p}}(t)+\mathbf{f}(s(t))+\mathbf{C}_{t}(t)
𝐂t​(t)=γ​𝐑​(t)​𝐩˙​(t)​θ​(t)​exp⁡(−β​θ​(t))θ⁡(t)=arccos⁡({𝐨−𝐩⁡(t)}T​𝐩˙​(t)‖𝐨−𝐩⁡(t)‖​‖𝐩˙​(t)‖)\displaystyle\begin{split}\mathbf{C}_{t}(t)&=\gamma\mathbf{R}(t)\dot{\mathbf{p}}(t)\theta(t)\exp(-\beta\theta(t))\\ \theta(t)&=\arccos{\bigg(\frac{\{\mathbf{o}-\mathbf{p}(t)\}^{\text{T}}\dot{\mathbf{p}}(t)}{||\mathbf{o}-\mathbf{p}(t)||~||\dot{\mathbf{p}}(t)||}\bigg)}\end{split} (17)
𝐑⁡(t)\displaystyle\mathbf{R}(t) =𝐈3+sin⁡(±π2)​[𝐫^​(t)]+{1−cos⁡(±π2)}​[𝐫^​(t)]2\displaystyle=\mathbf{I}_{3}+\sin\big(\pm\frac{\pi}{2}\big)[\hat{\mathbf{r}}(t)]+\Big\{1-\cos\big(\pm\frac{\pi}{2}\big)\Big\}[\hat{\mathbf{r}}(t)]^{2}
=𝐈3±[𝐫^​(t)]+[𝐫^​(t)]2\displaystyle=\mathbf{I}_{3}\pm[\hat{\mathbf{r}}(t)]+[\hat{\mathbf{r}}(t)]^{2}

3.5.2 Elementary Dynamic Actions

A repulsive impedance is added to the goal-attracting task-space impedance. The resulting force changes the actual motion without explicitly constructing a new collision-free path.

Equation 18
𝐙p,1​(t)\displaystyle\mathbf{Z}_{p,1}(t) =𝐊p{𝐩0(t)−𝐩(t))+𝐁p{𝐩˙0(t)−𝐩˙(t)}\displaystyle=\mathbf{K}_{p}\{\mathbf{p}_{0}(t)-\mathbf{p}(t))+\mathbf{B}_{p}\{\dot{\mathbf{p}}_{0}(t)-\dot{\mathbf{p}}(t)\}
𝐙p,2​(t)\displaystyle\mathbf{Z}_{p,2}(t) =−k‖𝐨−𝐩⁡(t)‖n​(𝐨−𝐩​(t))\displaystyle=-\frac{k}{||\mathbf{o}-\mathbf{p}(t)||^{n}}(\mathbf{o}-\mathbf{p}(t))
𝝉i​n​(t)\displaystyle\bm{\tau}_{in}(t) =𝐉​(𝐪)T​{𝐙p,1​(t)+𝐙p,2​(t)}\displaystyle=\mathbf{J}(\mathbf{q})^{\text{T}}\{\mathbf{Z}_{p,1}(t)+\mathbf{Z}_{p,2}(t)\} (18)

3.5.3 Simulation Example

Both implementations avoid the illustrated obstacle. The EDA example needs a manually chosen 2 cm lateral offset to select a useful direction; without it the arm can stall or collide with a link. The result therefore does not establish complete obstacle avoidance from arbitrary starts.

Original figure
Fig. 7: A goal-directed discrete movement with obstacle avoidance for Dynamic Movement Primitives (DMPs, blue) and Elementary Dynamic Actions (EDAs, orange) (Section 3.5.3). (Left) Trajectory of the end-effector position. Initial end-effector position 𝐩i\mathbf{p}_{i} and goal location 𝐠\mathbf{g} are depicted as black circle markers. (Right) Time-lapse of the movement for DMP (Top, blue) and EDA (Bottom, orange). Dotted lines are a minimum-jerk trajectory. Obstacle location 𝐨=[0.0,1.14]\mathbf{o}=[0.0,1.14]m. Parameters of DMPs: γ=300\gamma=300, β=−3\beta=-3 (Equation (17)). Parameters of EDAs: k=0.1k=0.1N/m5, n=6n=6 (Equation (18)). Other parameters are identical to those in Figure 4.

3.6 Rhythmic Movement

Rhythmic examples use sinusoidal joint motion and a circular end-effector path.

3.6.1 Dynamic Movement Primitives

Periodic DMP phase and forcing functions generate the reference kinematics, followed by the same joint/task mapping and inverse dynamics used earlier.

3.6.2 Elementary Dynamic Actions

A periodic virtual trajectory drives the existing joint-space or task-space impedance controller.

3.6.3 Simulation Example

Both methods generate periodic motion. The ideal-model DMP controller reproduces the reference closely, while EDAs exhibit tracking error but avoid inverse dynamics and explicit inverse kinematics in these examples.

Equations 19–20
𝐪d​e​s​(t),𝐪0​(t)=𝐪i+𝐪A​sin⁡(ω0​t)\mathbf{q}_{des}(t),\mathbf{q}_{0}(t)=\mathbf{q}_{i}+\mathbf{q}_{A}\sin(\omega_{0}t) (19)
𝐩d​e​s​(t),𝐩0​(t)=𝐜+[r0​cos⁡(ω0​t),r0​sin⁡(ω0​t)]\mathbf{p}_{des}(t),\mathbf{p}_{0}(t)=\mathbf{c}+[r_{0}\cos(\omega_{0}t),r_{0}\sin(\omega_{0}t)] (20)
Refer to caption
Fig. 8: Rhythmic movements in (A, B, C, D) joint-space and (E, F) task-space for Dynamic Movement Primitives (DMPs, blue) and Elementary Dynamic Actions (EDAs, orange) (Section 3.6.3). (A, B) Rhythmic movement in joint-space and its (C, D) joint trajectories. (C, D) The black dashed lines which are perfectly overlapped with DMPs (blue lines) represent a sinusoidal trajectory with parameters: 𝐪i=[0.5,0.5]\mathbf{q}_{i}=[0.5,0.5] rad, 𝐪A=[0.1,0.3]\mathbf{q}_{A}=[0.1,0.3] rad, ω0=π\omega_{0}=\pi rad/s (Equation (19)). Parameters of rhythmic DMP: αz=10\alpha_{z}=10, βz=2.5\beta_{z}=2.5, τ=1.0\tau=1.0, r=1.0r=1.0, N=40N=40, P=100P=100, ci=2​π​(i−1)/Nc_{i}=2\pi(i-1)/N, hi=Nh_{i}=N for i∈[1,2,⋯,N]i\in[1,2,\cdots,N]. Parameters of EDA are identical to those in Figure 3. (E, F) Rhythmic movement in task-space. The black dashed lines represent a circular trajectory with parameters: 𝐜=[0,1.4142]\mathbf{c}=[0,1.4142]m, r0=0.5r_{0}=0.5m, ω0=π\omega_{0}=\pi rad/s (Equation (20)). Parameters of EDAs: 𝐊p=90​𝐈2\mathbf{K}_{p}=90\mathbf{I}_{2} N/m, 𝐁p=60​𝐈2\mathbf{B}_{p}=60\mathbf{I}_{2} N⋅\cdots/m. Higher stiffness and damping values than Figure 4 were chosen to reduce the tracking error. Parameters of rhythmic DMP are identical to those in (A, B). Consistent with Figure 3 and Figure 4, for DMP, perfect tracking was achieved while for EDA, a non-negligible tracking error existed.

3.7 Combination of Discrete and Rhythmic Movements

The next examples shift the center of an ongoing rhythmic movement.

Refer to caption
Fig. 9: A combination of discrete and rhythmic movements in (A-F) joint-space and (G-J) task-space for Dynamic Movement Primitives (DMPs, blue) and Elementary Dynamic Actions (EDAs, orange) (Section 3.7.3). (C, E, I) Black filled lines represent the discrete change of 𝐠0\mathbf{g}_{0} of DMP. Black dotted lines represent 𝐠⁡(t)\mathbf{g}(t), which follows a response of a (critically damped) second-order linear system. (D, F, J) Black dotted lines represent 𝐪0​(t)\mathbf{q}_{0}(t) and 𝐩0​(t)\mathbf{p}_{0}(t) for joint-space and task-space, respectively. (A-F) Parameters of the discrete and rhythmic movements in joint-space: ω0=π\omega_{0}=\pi rad/s, 𝐪A=[0.1,0.3]\mathbf{q}_{A}=[0.1,0.3] rad, 𝐪i=[0.5,0.5]\mathbf{q}_{i}=[0.5,0.5] rad, 𝐪f=[1.5,1.5]\mathbf{q}_{f}=[1.5,1.5] rad. (G-J) Parameters of the discrete and rhythmic movements in task-space: ω0=π\omega_{0}=\pi rad/s, r0=0.3r_{0}=0.3m, 𝐩i=[−0.47,0.9]\mathbf{p}_{i}=[-0.47,0.9]m, 𝐩f=[0.53,0.9]\mathbf{p}_{f}=[0.53,0.9]m. In joint-space (respectively task-space), multiple discrete movements in opposite directions are generated by switching the values of 𝐪i\mathbf{q}_{i} (respectively 𝐩i\mathbf{p}_{i}) and 𝐪f\mathbf{q}_{f} (respectively 𝐩f\mathbf{p}_{f}), with time offset to​f​f=3.5+5​it_{off}=3.5+5i, where ii is the number of discrete movements. Parameters of discrete DMPs in both joint-space and task-space: τ2=1.0\tau_{2}=1.0, αz,2=10\alpha_{z,2}=10, βz,2=2.5\beta_{z,2}=2.5. Other parameters of DMP are identical to those in Figure 8. Parameters of EDA are identical to those in Figure 8. For EDA, given a single impedance operator, discrete and rhythmic movements can be directly combined at the level of the virtual trajectory.

3.7.1 Dynamic Movement Primitives

A discrete system generates a smoothly changing goal for the rhythmic DMP. This couples the two types rather than adding their learned forcing weights directly.

Equation
τ12​𝐪¨​(t)=−αz,1​βz,1​{𝐪⁡(t)−𝐠⁡(t)}−αz,1​τ1​𝐪˙​(t)+𝐟⁡(s⁡(t))s⁡(t)=tτ1mod2πτ22​𝐠¨​(t)=−αz,2​βz,2​{𝐠⁡(t)−𝐠0}−αz,2​τ2​𝐠˙​(t)\displaystyle\begin{aligned} \tau_{1}^{2}\ddot{\mathbf{q}}(t)&=-\alpha_{z,1}\beta_{z,1}\{\mathbf{q}(t)-\mathbf{g}(t)\}-\alpha_{z,1}\tau_{1}\dot{\mathbf{q}}(t)+\mathbf{f}(s(t))\\ s(t)&=\frac{t}{\tau_{1}}~~~~~~~~~~~~~~~\mod 2\pi\\ \tau_{2}^{2}\ddot{\mathbf{g}}(t)&=-\alpha_{z,2}\beta_{z,2}\{\mathbf{g}(t)-\mathbf{g}_{0}\}-\alpha_{z,2}\tau_{2}\dot{\mathbf{g}}(t)\\ \end{aligned}

3.7.2 Elementary Dynamic Actions

A finite submovement is directly added to an oscillation to form the virtual trajectory, leaving the impedance controller unchanged.

Equation
𝐪0​(t)=𝐪0,s​u​b​(t)+𝐪0,o​s​c​(t)\mathbf{q}_{0}(t)=\mathbf{q}_{0,sub}(t)+\mathbf{q}_{0,osc}(t)

3.7.3 Simulation Example

Both constructions produce combined discrete and rhythmic motion. The selected DMP goal change follows a second-order system, while the EDA construction allows a freely chosen submovement profile. This illustrates the simplicity of virtual-trajectory superposition for the chosen formulations.

Equations
𝐪0​(t)\displaystyle\mathbf{q}_{0}(t) =𝐪0,s​u​b​(t)+𝐪0,o​s​c​(t)\displaystyle=\mathbf{q}_{0,sub}(t)+\mathbf{q}_{0,osc}(t)
𝐪0,s​u​b​(t)\displaystyle\mathbf{q}_{0,sub}(t) ={𝟎 0≤t<to​f​f𝐪i+(𝐪f−𝐪i)​fM​J​T​(t) to​f​f≤t<T+to​f​f𝐪f T+to​f​f≤t\displaystyle=\begin{cases}\mathbf{0}&\text{ $0\leq t<t_{off}$}\\ \mathbf{q}_{i}+(\mathbf{q}_{f}-\mathbf{q}_{i})f_{MJT}(t)&\text{ $t_{off}\leq t<T+t_{off}$}\\ \mathbf{q}_{f}&\text{ $T+t_{off}\leq t$}\end{cases}
𝐪0,o​s​c​(t)\displaystyle\mathbf{q}_{0,osc}(t) =𝐪A​sin⁡(ω0​t)\displaystyle=\mathbf{q}_{A}\sin(\omega_{0}t)
fM​J​T​(t)\displaystyle f_{MJT}(t) =10​(t−to​f​fT)3−15​(t−to​f​fT)4+6​(t−to​f​fT)5\displaystyle=10\Big(\frac{t-t_{off}}{T}\Big)^{3}-15\Big(\frac{t-t_{off}}{T}\Big)^{4}+6\Big(\frac{t-t_{off}}{T}\Big)^{5}
𝐩0​(t)\displaystyle\mathbf{p}_{0}(t) =𝐩0,s​u​b​(t)+𝐩0,o​s​c​(t)\displaystyle=\mathbf{p}_{0,sub}(t)+\mathbf{p}_{0,osc}(t)
𝐩0,s​u​b​(t)\displaystyle\mathbf{p}_{0,sub}(t) ={𝟎 0≤t<to​f​f𝐩i+(𝐩f−𝐩i)​fM​J​T,o​f​f​(t) to​f​f≤t<T+to​f​f𝐠 T+to​f​f≤t\displaystyle=\begin{cases}\mathbf{0}&\text{ $0\leq t<t_{off}$}\\ \mathbf{p}_{i}+(\mathbf{p}_{f}-\mathbf{p}_{i})f_{MJT,off}(t)&\text{ $t_{off}\leq t<T+t_{off}$}\\ \mathbf{g}&\text{ $T+t_{off}\leq t$}\end{cases}
𝐩0,o​s​c​(t)\displaystyle\mathbf{p}_{0,osc}(t) =[r0​cos⁡(ω0​t),r0​sin⁡(ω0​t)]\displaystyle=[r_{0}\cos(\omega_{0}t),r_{0}\sin(\omega_{0}t)]

3.8 Sequence of Discrete Movements

A newly appearing target interrupts a discrete reach and requires a smooth redirection.

3.8.1 Dynamic Movement Primitives

Low-pass filtering the DMP goal avoids acceleration discontinuities. In the construction studied here, the initial learned motion is followed by a goal-change response governed by a third-order linear system.

Equations 21
τ2​𝐩¨​(t)=−αz​βz​{𝐩⁡(t)−𝐠⁡(t)}−αz​τ​𝐩˙​(t)+𝐟⁡(s⁡(t))τ​𝐠˙​(t)=αg​{𝐠0−𝐠⁡(t)}\displaystyle\begin{split}\tau^{2}\ddot{\mathbf{p}}(t)&=-\alpha_{z}\beta_{z}\{\mathbf{p}(t)-\mathbf{g}(t)\}-\alpha_{z}\tau\dot{\mathbf{p}}(t)+\mathbf{f}(s(t))\\ \tau\dot{\mathbf{g}}(t)&=\alpha_{g}\{\mathbf{g}_{0}-\mathbf{g}(t)\}\end{split} (21)
𝐠⁡(t)=𝐠n​e​w+(𝐠o​l​d−𝐠n​e​w)​exp⁡(−αgτ​t)\mathbf{g}(t)=\mathbf{g}_{new}+(\mathbf{g}_{old}-\mathbf{g}_{new})\exp\bigg({-\frac{\alpha_{g}}{\tau}t}\bigg)
Refer to caption
Fig. 10: A sequence of discrete movements for (A, B, D, E) Dynamic Movement Primitives (DMPs, blue) and (A, C, D, E) Elementary Dynamic Actions (EDAs, orange) (Section 3.8.3). (A) Trajectory of the end-effector position. (B, C) Time-lapse of the movement of the 2-DOF robot model. (D, E) Time vs. XX- and YY-coordinates of the end-effector trajectory. (A, B, C) The first movement headed toward 𝐠o​l​d\mathbf{g}_{old} (square grey marker), but at time t=tgt=t_{g} (D, E), the target switched location to 𝐠n​e​w\mathbf{g}_{new} (square black marker) which necessitated a second movement. Parameters of DMPs (for the first discrete movement): αg=1.0\alpha_{g}=1.0, τ=1.0\tau=1.0. Other parameters are identical to those in Figure 4. Parameters of EDAs: T1=1.0T_{1}=1.0s, T2=1.0T_{2}=1.0s, tg=0.5t_{g}=0.5s. 𝐠o​l​d=[−0.7,1.22]\mathbf{g}_{old}=[-0.7,1.22]m, 𝐠n​e​w=[0.8,1.72]\mathbf{g}_{new}=[0.8,1.72]m, 𝐩i=[0.0,0.52]\mathbf{p}_{i}=[0.0,0.52]m. Other parameters are identical to those in Figure 4. (D, E) For the new goal location 𝐠n​e​w\mathbf{g}_{new}, EDA achieved faster convergence than DMP. For DMP the second movement followed a response of a third-order linear system. On the other hand for EDA, the basis function of the second movement can be freely chosen. Hence, faster convergence to 𝐠n​e​w\mathbf{g}_{new} was achieved for EDA.

3.8.2 Elementary Dynamic Actions

A second submovement is superimposed on the first virtual trajectory, with amplitude measured from the previous goal to the new goal. Each submovement may use its own smooth basis.

Equation
𝐩˙0​(t)\displaystyle\dot{\mathbf{p}}_{0}(t) =∑i𝐯i​σ^i​(t−ti)\displaystyle=\sum_{i}\mathbf{v}_{i}\ \hat{\sigma}_{i}(t-t_{i})

3.8.3 Simulation Example

Both controllers converge to the newly supplied goal. The selected EDA profiles reach it faster in this example, and the first submovement need not be modified. This is a result of the compared constructions and tuning, not a universal speed advantage.

Equations
𝐠0​(t)={𝐠o​l​d 0≤t<tg 𝐠n​e​w tg≤t\displaystyle\begin{aligned} \mathbf{g}_{0}(t)&=\begin{cases}\mathbf{g}_{old}&\text{ $0\leq t<t_{g}$ }\\ \mathbf{g}_{new}&\text{ $t_{g}\leq t$ }\end{cases}\end{aligned}
𝐩0​(t)\displaystyle\mathbf{p}_{0}(t) =𝐩0,1​(t)+𝐩0,2​(t)\displaystyle=\mathbf{p}_{0,1}(t)+\mathbf{p}_{0,2}(t)
𝐩0,1​(t)\displaystyle\mathbf{p}_{0,1}(t) ={𝐩i+(𝐠o​l​d−𝐩i)​fM​J​T,1​(t) 0≤t<T1𝐠o​l​d T1≤t\displaystyle=\begin{cases}\mathbf{p}_{i}+(\mathbf{g}_{old}-\mathbf{p}_{i})f_{MJT,1}(t)&\text{ $0\leq t<T_{1}$}\\ \mathbf{g}_{old}&\text{ $T_{1}\leq t$}\end{cases}
𝐩0,2​(t)\displaystyle\mathbf{p}_{0,2}(t) ={𝟎 0≤t<tg(𝐠n​e​w−𝐠o​l​d)​fM​J​T,2​(t) tg≤t<tg+T2𝐠n​e​w−𝐠o​l​d tg+T2≤t\displaystyle=\begin{cases}\mathbf{0}&\text{ $0\leq t<t_{g}$}\\ (\mathbf{g}_{new}-\mathbf{g}_{old})f_{MJT,2}(t)&\text{ $t_{g}\leq t<t_{g}+T_{2}$}\\ \mathbf{g}_{new}-\mathbf{g}_{old}&\text{ $t_{g}+T_{2}\leq t$}\end{cases}
fM​J​T,1​(t)\displaystyle f_{MJT,1}(t) =10​(tT1)3−15​(tT1)4+6​(tT1)5\displaystyle=10\Big(\frac{t}{T_{1}}\Big)^{3}-15\Big(\frac{t}{T_{1}}\Big)^{4}+6\Big(\frac{t}{T_{1}}\Big)^{5}
fM​J​T,2​(t)\displaystyle f_{MJT,2}(t) =10​(t−tgT2)3−15​(t−tgT2)4+6​(t−tgT2)5\displaystyle=10\Big(\frac{t-t_{g}}{T_{2}}\Big)^{3}-15\Big(\frac{t-t_{g}}{T_{2}}\Big)^{4}+6\Big(\frac{t-t_{g}}{T_{2}}\Big)^{5}

3.9 Managing Kinematic Redundancy

A redundant manipulator has more joint degrees of freedom than task coordinates, leaving additional joint motions compatible with one end-effector motion.

3.9.1 Dynamic Movement Primitives

The DMP trajectory drives a feedback controller using a Jacobian pseudoinverse and reference velocities/accelerations. This provides asymptotic tracking under the stated model assumptions. Null-space objectives can be added, but may complicate passivity; singularities still require regularization.

Equations
𝐩˙r​(t)\displaystyle\dot{\mathbf{p}}_{r}(t) =𝐩˙d​e​s​(t)+𝚲1​(𝐩d​e​s​(t)−𝐩⁡(t))\displaystyle=\dot{\mathbf{p}}_{des}(t)+\bm{\Lambda}_{1}(\mathbf{p}_{des}(t)-\mathbf{p}(t))
𝐪˙r​(t)\displaystyle\dot{\mathbf{q}}_{r}(t) =𝐉​(𝐪)+​𝐩˙r​(t)=𝐉​(𝐪)+​{𝐩˙d​e​s​(t)+𝚲1​(𝐩d​e​s​(t)−𝐩⁡(t))}\displaystyle=\mathbf{J}(\mathbf{q})^{+}\dot{\mathbf{p}}_{r}(t)=\mathbf{J}(\mathbf{q})^{+}\big\{\dot{\mathbf{p}}_{des}(t)+\bm{\Lambda}_{1}(\mathbf{p}_{des}(t)-\mathbf{p}(t))\big\}
𝐩¨r​(t)\displaystyle\ddot{\mathbf{p}}_{r}(t) =𝐩¨d​e​s​(t)+𝚲1​(𝐩˙d​e​s​(t)−𝐩˙​(t))\displaystyle=\ddot{\mathbf{p}}_{des}(t)+\bm{\Lambda}_{1}(\dot{\mathbf{p}}_{des}(t)-\dot{\mathbf{p}}(t))
𝐪¨r​(t)\displaystyle\ddot{\mathbf{q}}_{r}(t) =𝐉​(𝐪)+​(𝐩¨r​(t)−𝐉˙​(𝐪)​𝐪˙​(t))\displaystyle=\mathbf{J}(\mathbf{q})^{+}(\ddot{\mathbf{p}}_{r}(t)-\mathbf{\dot{J}}(\mathbf{q})\dot{\mathbf{q}}(t))
=𝐉​(𝐪)+​{𝐩¨d​e​s​(t)+𝚲1​[𝐩˙d​e​s​(t)−𝐩˙​(t)]−𝐉˙​(𝐪)​𝐪˙​(t)}\displaystyle=\mathbf{J}(\mathbf{q})^{+}\big\{\ddot{\mathbf{p}}_{des}(t)+\bm{\Lambda}_{1}[\dot{\mathbf{p}}_{des}(t)-\dot{\mathbf{p}}(t)]-\mathbf{\dot{J}}(\mathbf{q})\dot{\mathbf{q}}(t)\big\}
𝝉i​n​(t)=𝐌⁡(𝐪)​𝐪¨r​(t)+𝐂⁡(𝐪,𝐪˙)​𝐪˙r​(t)−𝚲2​{𝐪˙​(t)−𝐪˙r​(t)}\bm{\tau}_{in}(t)=\mathbf{M}(\mathbf{q})\ddot{\mathbf{q}}_{r}(t)+\mathbf{C}(\mathbf{q},\dot{\mathbf{q}})\dot{\mathbf{q}}_{r}(t)-\bm{\Lambda}_{2}\{\dot{\mathbf{q}}(t)-\dot{\mathbf{q}}_{r}(t)\}
Refer to caption
Fig. 11: Managing kinematic redundancy using Dynamic Movement Primitives (DMPs, blue) and Elementary Dynamic Actions (EDAs, orange) (Section 3.9.3). (A, B, C) A single of discrete movement. (D, E, F) A sequence of discrete movements. The first movement headed toward 𝐠o​l​d\mathbf{g}_{old} (square grey marker), but at time t=tgt=t_{g}, target switched location to 𝐠n​e​w\mathbf{g}_{new} (square black marker) which necessitated a second movement. (A, B) The dashed black lines are a minimum-jerk trajectory. (C, F) End-effector trajectories. (A, B, C) Parameters of the minimum-jerk trajectory: 𝐩i=[0,3]\mathbf{p}_{i}=[0,3]m, 𝐠=[3,3]\mathbf{g}=[3,3]m, T=2.0T=2.0s. (D, E, F) Parameters of the minimum-jerk trajectories: 𝐩i=[−1.62,0.76]\mathbf{p}_{i}=[-1.62,0.76]m, 𝐠o​l​d=[−3.62,1.76]\mathbf{g}_{old}=[-3.62,1.76]m, 𝐠n​e​w=[2.38,3.26]\mathbf{g}_{new}=[2.38,3.26]m, T1=2.0T_{1}=2.0s, T2=3.0T_{2}=3.0s, tg=1.0t_{g}=1.0. Parameters of DMPs: 𝚲1=80​𝐈2\bm{\Lambda}_{1}=80\mathbf{I}_{2}, 𝚲2=100​𝐈5\bm{\Lambda}_{2}=100\mathbf{I}_{5}, τ=T\tau=T. Other parameters are identical to those in Figure 3. Parameters of EDAs: 𝐊p=300​𝐈2\mathbf{K}_{p}=300\mathbf{I}_{2} N/m, 𝐁p=100​𝐈2\mathbf{B}_{p}=100\mathbf{I}_{2} N⋅\cdots/m, 𝐁q=30​𝐈5\mathbf{B}_{q}=30\mathbf{I}_{5} N⋅\cdotm⋅\cdots/rad.

3.9.2 Elementary Dynamic Actions

Task-space impedance is combined with joint damping to suppress otherwise uncontrolled joint motion. No explicit inverse kinematics is needed. Repetitive motion can still drift in joint space; adding stiffness may control drift but can conflict with the task objective. Superposition must therefore be designed with those tradeoffs in mind.

Equation 22
𝐙p​(t)=𝐊p​{𝐩0​(t)−𝐩⁡(t)}+𝐁p​{𝐩˙0​(t)−𝐩˙​(t)}𝐙q​(t)=−𝐁q​𝐪˙​(t)𝝉i​n​(t)=𝐉​(𝐪)T​𝐙p​(t)+𝐙q​(t)\displaystyle\begin{split}\mathbf{Z}_{p}(t)&=\mathbf{K}_{p}\{\mathbf{p}_{0}(t)-\mathbf{p}(t)\}+\mathbf{B}_{p}\{\dot{\mathbf{p}}_{0}(t)-\dot{\mathbf{p}}(t)\}\\ \mathbf{Z}_{q}(t)&=-\mathbf{B}_{q}\dot{\mathbf{q}}(t)\\ \bm{\tau}_{in}(t)&=\mathbf{J}(\mathbf{q})^{\text{T}}\mathbf{Z}_{p}(t)+\mathbf{Z}_{q}(t)\end{split} (22)

3.9.3 Simulation Example

A five-link planar arm performs single and sequential reaches with both controllers. The EDA implementation reuses the previous task impedance and adds joint damping, illustrating modular reuse. The DMP implementation uses its generated reference with the more elaborate feedback mapping.

4 Discussion

The discussion separates shared modular principles from differences in the chosen trajectory and interaction-control implementations.

4.1 Similarities between the Two Approaches

Both approaches reduce motion design to reusable modules and parameters. Shared DMP phase coordinates many joints, while EDA submovements and impedance components simplify combined tasks. Their modularity concerns different mathematical objects, so reuse does not imply identical composition rules.

4.2 Differences between the Two Approaches

The main tradeoffs concern model requirements, tracking, interaction dynamics, singularities, and the level at which modules combine.

4.2.1 The Need for an Inverse Dynamics Model

Inverse dynamics is central to the torque-tracking DMP implementation examined here, while joint-position-controlled robots can execute DMP references through existing servos. Position control alone, however, does not expose the interaction dynamics needed for compliant contact. The comparison should separate the trajectory generator from the robot’s low-level actuation and controller.

4.2.2 Tracking Control

DMPs provide a convenient learned reference for accurate tracking under ideal conditions. EDAs prioritize the impedance relation and need suitable virtual trajectories, impedance choices, or optimization if precise motion tracking is also required.

4.2.3 Contact and Physical Interaction

EDAs treat physical interaction dynamics as an explicit primitive; DMP implementations can add comparable impedance feedback. A PD equation and a first-order joint impedance coincide only with ideal torque sources, and general impedance control can act at other interaction ports.

4.2.4 Managing Kinematic Singularity and Redundancy

Jacobian-transpose impedance commands avoid singular inverses, but do not remove all motion-design concerns. Pseudoinverse-based DMP tracking needs singularity handling and can lose passivity with added null-space projections. Stability claims depend on the complete controller, not just the primitive’s label.

4.2.5 Modularity in Robot Control

EDA virtual trajectories and compatible impedance components admit direct superposition. DMP weights are reusable, but combining distinct motions requires an explicit coordination rule rather than simply merging weight vectors. This distinction explains the sequencing and mixed-motion examples.

4.3 Combining the Best of Both Approaches

Hybrid designs can use DMPs to learn rich virtual trajectories and impedance control to regulate interaction, or add impedance feedback to inverse-dynamics tracking. The paper points to prior variable-impedance learning as an example rather than introducing a newly evaluated hybrid algorithm.

5 Conclusion

The eight simulations clarify tradeoffs between learned trajectory generation and explicit interaction control. They support informed controller design under known assumptions, but provide no physical validation, broad performance ranking, or unconditional safety guarantee.

Acknowledgements

Funding

The work acknowledges MIT/SUSTech research support and a MathWorks fellowship for Moses C. Nah.

Declaration of Conflicting Interests

The authors declare no conflicting interests.

Concept demonstration

Open image ↗