Literature
Display
← All papers

Passive Obstacle Aware Control to Follow Desired Velocities

Lukas Huber · Thibaud Trinca · Jean-Jacques Slotine · Aude Billard

· IEEE Robotics and Automation Letters ↗

Robotics8 min text · CondensedOriginal paper ↗
Contents

At a glance

This controller follows a desired collision-free velocity field while changing its damping near obstacles. It stays compliant in free space, rejects motion toward nearby surfaces, and better resists disturbances in simulation and on a seven-joint Franka arm. The collision theorem has restricted geometry and impact assumptions, and the passivity analysis explicitly allows a bounded nonpassive velocity region. This edition preserves the requested first arXiv version; the publication link points to the later journal article.

I Introduction

A collision-free velocity command does not guarantee that an inertial, compliant robot will follow it after a push. The controller therefore uses both the desired motion and local obstacle geometry to decide which disturbances to resist. Its main tradeoff is compliance away from obstacles versus stronger corrective forces near them.

Refer to caption

Fig. 1: The proposed passive obstacle-aware controller lets the robot absorb external disturbances while ensuring collision avoidance. While tipping over the closed pasta box on this dinner table setup might be acceptable. Yet, the delicate wine glasses demand careful handling to prevent breakage.

I-A Related Work

The background covers interaction-force control and reactive obstacle avoidance.

I-A 1 Force Control

Impedance and damping controllers map motion errors to interaction forces. Variable parameters improve adaptability but need an energy or stability analysis. Previous passive velocity-following designs prioritize the motion direction and can remain too compliant toward an obstacle; energy-tank methods add another way to bound energy but may limit available behavior.

I-A 2 Obstacle Avoidance

Potential fields, navigation functions, geometric motion policies, and velocity modulation provide reactive motion around obstacles. This work assumes an appropriate velocity field is already available and changes the force controller that tracks it. The failure it targets is a disturbance pushing a compliant robot away from that field and into an obstacle.

I-B Contribution

The contributions are an obstacle-sensitive damping matrix, a passivity-region and boundedness analysis, a conditional collision-avoidance argument, a discrete-time parameter analysis, and simulation/hardware comparisons.

Original figure
Fig. 2: The desired velocity 𝒇b​(𝝃)\boldsymbol{f}^{b}(\boldsymbol{\xi}) can result from a learned velocity field or pointing towards a desired attractor 𝝃a\boldsymbol{\xi}^{a}. The desired velocity is used to evaluate the obstacle avoidance velocity 𝒇⁡(𝝃)\boldsymbol{f}(\boldsymbol{\xi}), fed into the force controller to obtain the control force 𝝉c\boldsymbol{\tau}_{c}. In order to achieve collision avoidance, the distance function Γo​(𝝃)\Gamma_{o}(\boldsymbol{\xi}), the normal direction 𝒏o​(𝝃)\boldsymbol{n}_{o}(\boldsymbol{\xi}), and the reference direction 𝒓o​(𝝃)\boldsymbol{r}_{o}(\boldsymbol{\xi}) are evaluated for each obstacle o=1..Nobso=1..N^{\mathrm{{obs}}}.

II Preliminaries

The state can represent joint or Cartesian position in at least two dimensions. A continuous vector field supplies desired velocity; the physical system has second-order dynamics.

II-A Obstacle Avoidance

Velocity modulation reduces motion in an obstacle’s reference direction and increases tangential motion. A distance function equals one at the surface and grows outside it. The construction assumes star-shaped obstacles with a suitable internal reference point. This geometric layer supplies the collision-free desired field used by the damping controller.

Equations 1–3
𝒇⁡(𝝃)=E​(𝝃)​diag​(λr,λe,…,λe)​E​(𝝃)−1​𝒇b​(𝝃)withE​(𝝃)=[r​(𝝃)​e1​(𝝃)​…​ed−1​(𝝃)]\begin{split}&\boldsymbol{f}(\boldsymbol{\xi})=\textbf{E}(\boldsymbol{\xi})\text{diag}\left(\lambda^{r},\lambda^{e},...,\lambda^{e}\right)\textbf{E}(\boldsymbol{\xi})^{-1}\boldsymbol{f}^{b}(\boldsymbol{\xi})\\ &\text{with}\quad\textbf{E}(\boldsymbol{\xi})=\left[\textbf{r}(\boldsymbol{\xi})\ \textbf{e}_{1}(\boldsymbol{\xi})\ ...\ \textbf{e}_{d-1}(\boldsymbol{\xi})\right]\end{split} (1)
λr​(𝝃)=1−1/Γ⁡(𝝃),λe​(𝝃)=1+1/Γ⁡(𝝃)\begin{split}\lambda^{r}(\boldsymbol{\xi})=1-1/\Gamma(\boldsymbol{\xi}),\quad\lambda^{e}(\boldsymbol{\xi})=1+1/\Gamma(\boldsymbol{\xi})\end{split} (2)
Γ⁡(𝝃)=1+‖𝝃−𝝃b‖/R0with𝝃b=b⁡(𝝃−𝝃r)+𝝃r\Gamma(\boldsymbol{\xi})=1+\|\boldsymbol{\xi}-\boldsymbol{\xi}^{b}\|/R^{0}\quad\text{with}\quad\boldsymbol{\xi}^{b}=b(\boldsymbol{\xi}-\boldsymbol{\xi}^{r})+\boldsymbol{\xi}^{r} (3)

Original figure

Fig. 3: The Γ\Gamma-field is defined individually for each of the obstacles. At each position 𝝃\boldsymbol{\xi}, we can evaluate the surface normal 𝒏⁡(𝝃)\boldsymbol{n}(\boldsymbol{\xi}). The velocity 𝒇⁡(𝝃)\boldsymbol{f}(\boldsymbol{\xi}) (gray) avoids collision with the obstacles and converges towards the attractor (star).

II-B Force Control

Rigid-body dynamics and a gravity-compensated damping law connect desired velocity to control force.

II-B 1 Rigid Body Dynamics

The model includes inertia, Coriolis effects, gravity, applied control, and external disturbance force. The energy analysis uses the standard skew-symmetry property of the rigid-body inertia and Coriolis terms.

Equation 4
ℳ⁡(𝝃)​𝝃¨+𝒞⁡(𝝃,𝝃˙)​𝝃˙+𝒈⁡(𝝃)=𝝉𝒄+𝝉𝒆\mathcal{M}(\boldsymbol{\xi})\boldsymbol{\ddot{\xi}}+\mathcal{C}(\boldsymbol{\xi},\boldsymbol{\dot{\xi}})\boldsymbol{\dot{\xi}}+\boldsymbol{g}(\boldsymbol{\xi})=\boldsymbol{\tau_{c}}+\boldsymbol{\tau_{e}} (4)

II-B 2 Damping Controller

The baseline applies gravity compensation plus D times desired-minus-actual velocity. An orthonormal basis aligns the largest damping direction with the desired motion, while smaller transverse values permit compliance. All damping eigenvalues are positive.

Equations 5–7
𝝉𝒄=𝒈⁡(𝝃)+𝒟⁡(𝝃)​(𝒇⁡(𝝃)−𝝃˙)\boldsymbol{\tau_{c}}=\boldsymbol{g}(\boldsymbol{\xi})+\mathcal{D}(\boldsymbol{\xi})\left(\boldsymbol{f}(\boldsymbol{\xi})-\boldsymbol{\dot{\xi}}\right) (5)
𝒟⁡(𝝃)=𝒬⁡(𝝃)​𝒮​(𝝃)​𝒬​(𝝃)−1\mathcal{D}(\boldsymbol{\xi})=\mathcal{Q}(\boldsymbol{\xi})\mathcal{S}(\boldsymbol{\xi})\mathcal{Q}(\boldsymbol{\xi})^{-1} (6)
𝒒1​(𝝃)=𝒒1f​(𝝃)=𝒇⁡(𝝃)/∥𝒇⁡(𝝃)∥\boldsymbol{q}_{1}(\boldsymbol{\xi})=\boldsymbol{q}_{1}^{f}(\boldsymbol{\xi})=\boldsymbol{f}({\boldsymbol{\xi}})/\lVert\boldsymbol{f}({\boldsymbol{\xi}})\rVert (7)

II-C Stability Analysis

Passivity compares change in stored energy with supplied input–output power. It is useful for interaction analysis, but a robot accelerating toward a nonzero commanded velocity can inject kinetic energy without an external push. The later result therefore identifies where the passivity inequality holds rather than asserting it at every velocity.

Equation 8
∫0t∗s⁡(u⁡(t),y⁡(t))​𝑑τ≥S⁡(t∗)−S⁡(0)\int_{0}^{t^{*}}s\left(u(t),y(t)\right)d\tau\geq S(t^{*})-S(0) (8)

II-D Problem Statement

The desired velocity must be continuous, bounded, and directed tangentially or outward at obstacle boundaries. These conditions apply to the commanded first-order field. They do not by themselves guarantee that the disturbed second-order robot remains outside an obstacle.

III Obstacle Aware Passivity

The controller blends velocity-following damping with obstacle-oriented damping. A danger weight increases near a surface and depends on the averaged obstacle normal. Positive-definite components and a weight between zero and one keep the blended damping positive definite.

Equations 9–10
𝒟⁡(𝝃)=(1−w⁡(𝝃))​𝒟f​(𝝃)+w⁡(𝝃)​𝒟o​(𝝃)\mathcal{D}(\boldsymbol{\xi})=\left(1-w(\boldsymbol{\xi})\right){\mathcal{D}^{f}}(\boldsymbol{\xi})+w(\boldsymbol{\xi}){\mathcal{D}^{\mathrm{o}}}(\boldsymbol{\xi}) (9)
w⁡(𝝃)=max⁡(0,Γcrit−Γ⁡(𝝃)Γcrit−1)​‖𝒏⁡(𝝃)‖withΓ⁡(𝝃)=mino=1..Nobs⁡Γo​(𝝃)\begin{split}w(\boldsymbol{\xi})=\max\left(0,\frac{\Gamma^{\mathrm{crit}}-\Gamma(\boldsymbol{\xi})}{\Gamma^{\mathrm{crit}}-1}\right)\|\boldsymbol{n}(\boldsymbol{\xi})\|\\ \text{with}\quad\Gamma(\boldsymbol{\xi})=\min_{o=1..N^{\mathrm{obs}}}\Gamma_{o}(\boldsymbol{\xi})\end{split} (10)
Original figure
Fig. 4: The damping matrix enforcing desired velocity following 𝒟f\mathcal{D}^{f} has the first basis vector 𝒒1f\boldsymbol{q}_{1}^{f} which points along the avoidance velocity 𝒇⁡(𝝃)\boldsymbol{f}(\boldsymbol{\xi}). The damping matrix to enforce collision avoidance 𝒟obs\mathcal{D}^{\mathrm{obs}} uses the normal 𝒏⁡(𝝃)\boldsymbol{n}(\boldsymbol{\xi}) to construct the first direction of the decomposition basis 𝒒1obs\boldsymbol{q}^{\mathrm{obs}}_{1}.

III-A Damping for Collision Repulsion

The obstacle component chooses a direction toward nearby surfaces, an orthogonal tracking direction, and corresponding damping strengths.

III-A 1 Normal Direction

Inverse-distance weights combine obstacle normals, emphasizing the closest surfaces. The result approaches the nearest obstacle’s normal at its boundary. Opposing obstacles can cancel the average, so narrow passages require an additional rule.

Equation 11
𝒏⁡(𝝃)=∑o=1No𝒏𝒐​(𝝃)​1/(Γo​(𝝃)−1)∑p=1No1/(Γp​(𝝃)−1)\boldsymbol{n}(\boldsymbol{\xi})=\sum_{o=1}^{N^{\mathrm{o}}}\boldsymbol{n_{o}}(\boldsymbol{\xi})\frac{1/(\Gamma_{o}(\boldsymbol{\xi})-1)}{\sum_{p=1}^{N^{\mathrm{o}}}1/(\Gamma_{p}(\boldsymbol{\xi})-1)} (11)

III-A 2 Decomposition Matrix

The first basis vector follows the averaged normal; the second follows the component of desired velocity orthogonal to it. Remaining vectors complete an orthonormal basis. Parallel normal/velocity directions permit an arbitrary transverse basis, with equal transverse damping used to avoid dependence on that choice.

Equations 12–13
𝒒1o​(𝝃)=𝒏⁡(𝝃)/∥𝒏⁡(𝝃)∥∀𝝃:∥𝒏⁡(𝝃)∥>0\boldsymbol{q}_{1}^{\mathrm{o}}(\boldsymbol{\xi})=\boldsymbol{n}(\boldsymbol{\xi})/\lVert\boldsymbol{n}(\boldsymbol{\xi})\rVert\quad\forall\,\boldsymbol{\xi}:\lVert\boldsymbol{n}(\boldsymbol{\xi})\rVert>0 (12)
𝒒2o=𝒒^2o‖𝒒^2o‖𝒒^2o=𝒒1f−𝒒1o​p∀𝝃:|p|<1\boldsymbol{q}_{2}^{\mathrm{o}}=\frac{\hat{\boldsymbol{q}}_{2}^{\mathrm{o}}}{\|\hat{\boldsymbol{q}}_{2}^{\mathrm{o}}\|}\quad\hat{\boldsymbol{q}}_{2}^{\mathrm{o}}=\boldsymbol{q}_{1}^{f}-\boldsymbol{q}_{1}^{\mathrm{o}}p\quad\forall\,\boldsymbol{\xi}:|p|<1 (13)

III-A 3 Damping Values

Large normal damping resists motion toward an obstacle, intermediate damping follows the desired velocity, and small remaining values preserve compliance. Interpolating transverse damping as the normal aligns with the velocity helps keep the resulting control force continuous.

Equations 14–15
𝒮do​(𝝃)={sod=1|p|​sc+(1−|p|)​sfd=2scd=3..N\mathcal{S}_{d}^{\mathrm{o}}(\boldsymbol{\xi})=\begin{cases}s^{\mathrm{o}}&d=1\\ |p|s^{c}+(1-|p|)s^{f}&d=2\\ s^{c}&d=3..N\end{cases} (14)
lim|p|→1𝒮d=𝒮e,d>2..N,e>2..N\lim_{|p|\rightarrow 1}\mathcal{S}_{d}=\mathcal{S}_{e},\quad d>2..N,\,e>2..N (15)

III-A 4 Damping Only Towards Obstacle

Strong normal damping is used when the velocity error requires correction away from the obstacle. Moving away from the surface can remain compliant. Although the selected damping coefficient switches, its multiplying error component is zero at the switch, preserving force continuity in the stated construction.

Equation 16
𝒮1o​(𝝃)={soif​(𝒇⁡(𝝃)−𝝃˙)T​𝒏​(𝝃)>0scotherwise\mathcal{S}_{1}^{\mathrm{o}}(\boldsymbol{\xi})=\begin{cases}s^{\mathrm{o}}&\text{if}\;\;\left(\boldsymbol{f}(\boldsymbol{\xi})-\boldsymbol{\dot{\xi}}\right)^{T}\boldsymbol{n}(\boldsymbol{\xi})>0\\ s^{\mathrm{c}}&\text{otherwise}\end{cases} (16)

III-B Damping for Velocity Preservation

The velocity-following component normally emphasizes the commanded direction. Near opposing surfaces, its eigenvalues are raised so that cancellation of the average normal does not leave a compliant direction toward either obstacle.

Equation 17
𝒮df=wp​so+{(1−wp)​sfd=1(1−wp)​ssd=2..Nwithwp=min⁡(1,‖𝒏⁡(𝝃)‖2+Δ​Γ2)andΔ​Γ=max⁡(Γcrit−Γ⁡(𝝃)Γcrit−1,0)\begin{split}&\mathcal{S}^{f}_{d}=w^{p}s^{\mathrm{o}}+\begin{cases}(1-w^{p})s^{f}&d=1\\ (1-w^{p})s^{s}&d=2..N\end{cases}\\ &\text{with}\quad w^{p}=\min\left(1,\|\boldsymbol{n}(\boldsymbol{\xi})\|^{2}+\Delta\Gamma^{2}\right)\\ &\text{and}\quad\Delta\Gamma=\max\left(\frac{\Gamma^{\mathrm{crit}}-\Gamma(\boldsymbol{\xi})}{\Gamma^{\mathrm{crit}}-1},0\right)\end{split} (17)

III-C Cluttered Environments

When two nearby surfaces have opposing normals, the obstacle-oriented weight tends to zero while the modified velocity component tends to high isotropic damping. This supplies disturbance resistance in narrow passages without requiring a well-defined averaged normal.

Equations 18–19
‖𝒏⁡(𝝃)‖→0andlimΓ⁡(𝝃)→1w⁡(𝝃)=0\|\boldsymbol{n}(\boldsymbol{\xi})\|\rightarrow 0\quad\text{and}\quad\lim_{\Gamma(\boldsymbol{\xi})\rightarrow 1}w(\boldsymbol{\xi})=0 (18)
limΓ→1,‖n‖→0𝒟⁡(𝝃)=𝒟S​(𝝃)+0=ℐ​so\lim_{\Gamma\rightarrow 1,\|n\|\rightarrow 0}\mathcal{D}(\boldsymbol{\xi})=\mathcal{D}^{S}(\boldsymbol{\xi})+0=\mathcal{I}s^{o} (19)

III-D Damping Parameter Design

The recommended ordering is obstacle damping at least as large as velocity-following damping, both much larger than the positive compliance value. Stronger damping improves rejection but raises forces and can destabilize a sampled implementation if chosen without regard to the update interval.

Equation 20
so≥sf≫sc>0s^{\mathrm{o}}\geq s^{f}\gg s^{c}>0 (20)

III-E Passivity Analysis

Using kinetic energy as storage, the derivative is external power minus a damping expression involving actual velocity and velocity error. For isotropic damping, the inequality is passive outside a circle in velocity space; a sufficient condition is speed at least as large as the commanded speed. Inside the identified region, the controller may add energy to accelerate toward the command.

An eigenbasis transformation gives the corresponding elliptical region for anisotropic damping. The paper then states a BIBO-stability result for bounded desired velocity and a continuous finite transformation. This is a velocity-boundedness claim under its model assumptions, not unconditional all-state passivity, target convergence, or collision avoidance. The complete original derivation is retained.

Equations 21–24
W⁡(𝝃,𝝃˙)=12​𝝃˙T​ℳ​(𝝃)​𝝃˙W(\boldsymbol{\xi},\boldsymbol{\dot{\xi}})=\frac{1}{2}\dot{\boldsymbol{\xi}}^{T}\mathcal{M}(\boldsymbol{\xi})\dot{\boldsymbol{\xi}} (21)
W˙​(𝝃,𝝃˙)=𝝃˙T​ℳ​(𝝃)​𝝃¨+12​𝝃˙T​ℳ˙​(𝝃)​𝝃˙\displaystyle\dot{W}(\boldsymbol{\xi},\boldsymbol{\dot{\xi}})=\boldsymbol{\dot{\xi}}^{T}\mathcal{M}(\boldsymbol{\xi})\boldsymbol{\ddot{\xi}}+\frac{1}{2}\boldsymbol{\dot{\xi}}^{T}\dot{\mathcal{M}}(\boldsymbol{\xi})\boldsymbol{\dot{\xi}}
=12​𝝃˙T​(ℳ˙​(𝝃)−2​𝒞​(𝝃))​𝝃˙−𝝃˙T​𝒟​(𝝃)​(𝝃˙−𝒇⁡(𝝃))+𝝃˙T​𝝉e\displaystyle=\frac{1}{2}\boldsymbol{\dot{\xi}}^{T}\left(\dot{\mathcal{M}}(\boldsymbol{\xi})-2\mathcal{C}(\boldsymbol{\xi})\right)\dot{\boldsymbol{\xi}}-\boldsymbol{\dot{\xi}}^{T}\mathcal{D}(\boldsymbol{\xi})\left(\boldsymbol{\dot{\xi}}-\boldsymbol{f}(\boldsymbol{\xi})\right)+\boldsymbol{\dot{\xi}}^{T}\boldsymbol{\tau}^{e}
=−𝝃˙T​𝒟​(𝝃)​(𝝃˙−𝐟⁡(𝝃))+𝝃˙T​𝝉e\displaystyle=-\boldsymbol{\dot{\xi}}^{T}\mathcal{D}(\boldsymbol{\xi})\left(\boldsymbol{\dot{\xi}}-\mathbf{f}(\boldsymbol{\xi})\right)+\boldsymbol{\dot{\xi}}^{T}\boldsymbol{\tau}^{e} (22)
𝒟⁡(𝝃)=𝒬⁡(𝝃)​𝒮​(𝝃)​𝒬​(𝝃)−1=𝒬⁡(𝝃)​ℐ​𝒬​(𝝃)−1=ℐ\mathcal{D}({\boldsymbol{\xi}})=\mathcal{Q}({\boldsymbol{\xi}})\mathcal{S}({\boldsymbol{\xi}})\mathcal{Q}({\boldsymbol{\xi}})^{-1}=\mathcal{Q}({\boldsymbol{\xi}})\mathcal{I}\mathcal{Q}({\boldsymbol{\xi}})^{-1}=\mathcal{I} (23)
𝝃˙T​𝒟​(𝝃)​(𝝃˙−𝒇⁡(𝝃))=𝝃˙T​Δ​𝒇≥0,Δ​𝒇=𝝃˙−𝒇⁡(𝝃)\dot{\boldsymbol{\boldsymbol{\xi}}}^{T}\mathcal{D}({\boldsymbol{\xi}})\left(\dot{\boldsymbol{\boldsymbol{\xi}}}-\boldsymbol{f}(\boldsymbol{\boldsymbol{\xi}})\right)=\dot{{\boldsymbol{\xi}}}^{T}\Delta\boldsymbol{f}\geq 0\;,\quad\Delta\boldsymbol{f}=\dot{{\boldsymbol{\xi}}}-\boldsymbol{f}({\boldsymbol{\xi}}) (24)
Original figure
Fig. 5: Analyzing the system in velocity-space yields that the system is passive if it has a velocity 𝝃˙\dot{\boldsymbol{\xi}} larger than the desired velocity 𝒇⁡(𝝃)\boldsymbol{f}(\boldsymbol{\xi}), i.e., outside the dashed circle. However, the system can be non-passive for small velocities when ⟨𝝃˙,Δ​𝒇⟩<0\left\langle{\dot{\boldsymbol{\xi}}},\,{\Delta\boldsymbol{f}}\right\rangle<0 (yellow circle).
Equations 25–29
W˙​(𝝃,𝝃˙)≤𝝃˙T​𝝉e∀𝝃:‖𝝃˙‖≥‖𝒇⁡(𝝃)‖\dot{W}({\boldsymbol{\xi}},\dot{{\boldsymbol{\xi}}})\leq\dot{{\boldsymbol{\xi}}}^{T}\boldsymbol{\tau}^{e}\quad\forall{\boldsymbol{\xi}}:\|\dot{{\boldsymbol{\xi}}}\|\geq\|\boldsymbol{f}({\boldsymbol{\xi}})\| (25)
∫0T‖𝝃˙‖​𝑑t=∫t∉𝒯n‖𝝃˙‖​𝑑t+∫t∈𝒯n‖𝝃˙‖​𝑑t≤Kp+vmax​Tn\begin{split}\int_{0}^{T}\left\|\dot{\boldsymbol{\xi}}\right\|\;dt&=\int_{t\notin\mathcal{T}_{n}}\left\|\dot{\boldsymbol{\xi}}\right\|\,dt+\int_{t\in\mathcal{T}_{n}}\left\|\dot{\boldsymbol{\xi}}\right\|\;dt\\ &\leq K_{p}+v^{\mathrm{max}}T_{n}\end{split} (26)
𝒗¯=𝒮⁡(𝝃)​𝒬​(𝝃)−1​𝝃˙​and​Δ​𝒇¯=𝒮⁡(𝝃)​𝒬​(𝝃)−1​Δ​𝒇\boldsymbol{\bar{v}}=\sqrt{\mathcal{S}({\boldsymbol{\xi}})}\mathcal{Q}({\boldsymbol{\xi}})^{-1}\dot{{\boldsymbol{\xi}}}\;\;\text{and}\;\;\bar{\Delta\boldsymbol{f}}=\sqrt{\mathcal{S}({\boldsymbol{\xi}})}\mathcal{Q}({\boldsymbol{\xi}})^{-1}\Delta\boldsymbol{f} (27)
𝝃˙T​𝒟​(𝝃)​Δ​𝒇=𝝃˙T​𝒬​(𝝃)​𝒮​(𝝃)​𝒬​(𝝃)−1​Δ​𝒇=𝒗¯T​Δ​𝒇¯\dot{\boldsymbol{\xi}}^{T}\mathcal{D}({\boldsymbol{\xi}})\Delta\boldsymbol{f}=\boldsymbol{\dot{\xi}}^{T}\mathcal{Q}({\boldsymbol{\xi}})\mathcal{S}({\boldsymbol{\xi}})\mathcal{Q}({\boldsymbol{\xi}})^{-1}\Delta\boldsymbol{f}=\boldsymbol{\bar{v}}^{T}\bar{\Delta\boldsymbol{f}} (28)
∫0T‖𝒗¯‖​𝑑t=∫t∉𝒯¯n‖𝒗¯‖​𝑑t+∫t∈𝒯¯n‖𝒗¯‖​𝑑t<Kp+vmax​T¯n​max⁡(eig​(𝒟))/min⁡(eig​(𝒟))\begin{split}&\int_{0}^{T}\left\|\boldsymbol{\bar{v}}\right\|\;dt=\int_{t\notin\bar{\mathcal{T}}_{n}}\left\|\boldsymbol{\bar{v}}\right\|\,dt+\int_{t\in\bar{\mathcal{T}}_{n}}\left\|\boldsymbol{\bar{v}}\right\|\;dt\\ &<K_{p}+v^{\mathrm{max}}\bar{T}_{n}{\max{\Bigl(\text{eig}\bigl(\mathcal{D}\bigr)\Bigr)}}/{\min{\Bigl(\text{eig}\bigl(\mathcal{D}\bigr)\Bigr)}}\end{split} (29)

Original figure

((a)) 𝒮1f>𝒮2f\mathcal{S}_{1}^{f}>\mathcal{S}_{2}^{f}, w≈0w\approx 0
Original figure
((b)) ⟨𝒇⁡(𝝃),𝒒1⟩≠‖𝒇⁡(𝝃)‖​‖𝒒2‖\left\langle{\boldsymbol{f}(\boldsymbol{\xi})},\,{\boldsymbol{q}_{1}}\right\rangle\neq\|\boldsymbol{f}(\boldsymbol{\xi})\|\;\|\boldsymbol{q}_{2}\|
Fig. 6: The stability is ensured even if the controller can temporarily accelerate the system to reach a velocity 𝝃˙\dot{\boldsymbol{\xi}}, which is faster than the desired velocity 𝒇⁡(𝝃)\boldsymbol{f}(\boldsymbol{\xi}) (white region). This happens when the eigenvalues of the damping matrix 𝒟⁡(𝝃)\mathcal{D}(\boldsymbol{\xi}) are not uniform (a) or the stretching vectors 𝒒1\boldsymbol{q}_{1} and 𝒒2\boldsymbol{q}_{2} are not orthogonal (b). In both cases, the region of non-passivity is elliptical (black circle).

IV Collision Avoidance

The collision lemma considers a point mass, a flat obstacle, a constant tangential desired velocity, and an impulsive disturbance. The derived displacement bound is proportional to impact speed times mass divided by damping. Sufficient clearance therefore limits the admissible impact magnitude.

Curved surfaces and varying velocity fields fall outside that argument and can still permit drift into obstacles. Force saturation further reduces the impact that can be rejected safely. The paper discusses repulsive-field additions for such cases; it does not establish collision avoidance under arbitrary force, geometry, or actuator limits.

Equations 30–35
𝒗I≈∫tItI+Δ​t𝝃¨​𝑑t≈∫tItI+Δ​tℳ−1​(𝝃)​𝝉e​𝑑t\displaystyle\boldsymbol{v}^{I}\approx\int_{t^{I}}^{t^{I}+\Delta t}\ddot{\boldsymbol{\xi}}dt\approx\int_{t^{I}}^{t^{I}+\Delta t}\mathcal{M}^{-1}(\boldsymbol{\xi})\boldsymbol{\tau}_{e}\,dt (30)
⟨𝒇⁡(𝝃),𝒏⁡(𝝃)⟩=0𝒇⁡(𝝃)=const.,𝒏⁡(𝝃)=const.\left\langle{\boldsymbol{f}(\boldsymbol{\xi})},\,{\boldsymbol{n}(\boldsymbol{\xi})}\right\rangle=0\qquad\boldsymbol{f}(\boldsymbol{\xi})=\text{const.}\,,\;\boldsymbol{n}(\boldsymbol{\xi})=\text{const.} (31)
|𝒏​(𝝃)T​{𝝃˙}t|=0∀Γ⁡(𝝃)=1\left|\boldsymbol{n}(\boldsymbol{\xi})^{T}\,\{\boldsymbol{\dot{\xi}}\}_{t}\right|=0\quad\forall\,\Gamma(\boldsymbol{\xi})=1 (32)
{𝝃˙}T=∫0T𝝃¨​𝑑t=∫0Tℳ−1​𝒟​(𝝃˙−𝒇⁡(𝝃))​𝑑t=ℳ−1​∫((1−w)​𝒟f+w​𝒟o)​(𝝃˙−𝒇⁡(𝝃))​𝑑t\begin{split}\{\boldsymbol{\dot{\xi}}\}_{T}&=\int_{0}^{T}\boldsymbol{\ddot{\xi}}\,dt=\int_{0}^{T}\mathcal{M}^{-1}\mathcal{D}\left(\boldsymbol{\dot{\xi}}-\boldsymbol{f}(\boldsymbol{\xi})\right)\,dt\\ &=\mathcal{M}^{-1}\int\left((1-w)\mathcal{D}^{f}+w\mathcal{D}^{o}\right)\left(\boldsymbol{\dot{\xi}}-\boldsymbol{f}(\boldsymbol{\xi})\right)dt\\ \end{split} (33)
0\displaystyle 0 =|𝒏​(𝝃)T​{𝝃˙}t|=1m​∫0T((1−w)​𝒏T​𝒟f​𝒏+w​so)​ξ˙​𝑑t\displaystyle=\left|\boldsymbol{n}(\boldsymbol{\xi})^{T}\,\{\boldsymbol{\dot{\xi}}\}_{t}\right|=\frac{1}{m}\int_{0}^{T}((1-w)\boldsymbol{n}^{T}\mathcal{D}^{f}\boldsymbol{n}+ws^{o})\,{\dot{\xi}}\,dt
<1m​∫0Tsf​ξ˙​𝑑t=sfm​(ξ−{ξ}0)+vI​𝑑t\displaystyle<\frac{1}{m}\int_{0}^{T}s^{f}\,{\dot{\xi}}\,dt=\frac{s^{f}}{m}(\xi-\{\xi\}_{0})+v^{I}\,dt (34)
‖{ξ}0−ξ‖≤‖vI‖​m/sf\|\{\xi\}_{0}-{\xi}\|\leq\|v^{I}\|{m}/{s^{\mathrm{f}}} (35)

Original figure

Fig. 7: A disturbance occurs of a point-agent at position 𝒑0\boldsymbol{p}^{0} with velocity after the impact of {𝝃˙}0=𝒗0+𝒗I\{\dot{\boldsymbol{\xi}}\}_{0}=\boldsymbol{v}^{0}+\boldsymbol{v}^{I}. A high damping in the direction of the obstacle in the presence of a constant velocity field (gray) ensures collision avoidance. Whereas different damping values sos^{\mathrm{o}} and optionally a maximum repulsion force 𝝉max\boldsymbol{\tau}^{\mathrm{max}} lead to different trajectories.

V Evaluation

The authors’ implementation is compared with a passive velocity-following controller in disturbance simulations and a physical robot experiment.

V-A Qualitative Repulse Rejection

Three simulated scenarios apply pushes toward a single obstacle, between opposing obstacles, and away from a surface. The obstacle-aware controller rejects the first two more strongly, while remaining similarly compliant for the outward push. The baseline collides in the shown inward-disturbance cases.

Refer to caption

Fig. 8: The desired velocity 𝒇⁡(𝝃)\boldsymbol{f}(\boldsymbol{\xi}) is represented in gray and serves as the input for the force controller. The mobile robot, initially positioned at three different locations, navigates safely towards the attractor (black star) even when confronted with external disturbances (purple arrows) while employing the obstacle-aware controller (blue trajectories). In contrast, the baseline controller (red) results in collisions when disturbances occur close to the robot.

V-B Noise Analysis

Noise experiments compare the minimum signed surface distance while following a bounded field through two obstacles. Position and velocity corruption test different sources of feedback error; the plots retain the complete tested noise ranges.

Refer to caption
((a)) Trajectories with a standard deviation of the velocity-noise of 1.0 m/s.
Original figure
((b)) The mean and variance (shaded) of the closest distances over 10 epochs.
Fig. 9: The agent navigates between two elongated obstacles (a) towards the attractor (black star). The agent’s velocity is subjected to white noise, with a mean of zero, and noise variances between 0.0 m/s0.0\text{\,}\mathrm{m}\mathrm{/}\mathrm{s} and 1.0 m/s1.0\text{\,}\mathrm{m}\mathrm{/}\mathrm{s}. The robot initiates its trajectory from the starting position 𝝃0=[−2.5,1.0]T\boldsymbol{\xi}_{0}=[-2.5,1.0]^{T}, with an initial velocity of zero. It aims to reach the attractor located at 𝝃a=[2.5,−1.0]T\boldsymbol{\xi}^{a}=[2.5,\,-1.0]^{T}.

V-B 1 Velocity Noise Resistance

With zero-mean velocity noise, stronger obstacle-directed damping maintains greater clearance than the baseline. Clearance decreases as noise rises, so the result is improved robustness over the tested range rather than immunity to noise.

V-B 2 Position Noise Resistance

With position noise, the obstacle-aware controller has positive mean minimum clearance for standard deviations below approximately 0.023 m in the reported setup. The baseline’s mean clearance is negative throughout those tests. A positive mean does not imply every trajectory avoids collision.

Refer to caption

((a)) Trajectories with a standard deviation of the position-noise of 0.0 m.
Original figure
((b)) Closest distances concerning different noise levels over ten epochs.
Fig. 10: The agent is navigating towards the attractor (black star) between two concave obstacles (a) while being subjected to white noise in its position. The position noise has a mean of zero and various noise variances ranging between 0.0 m0.0\text{\,}\mathrm{m} and 0.03 m0.03\text{\,}\mathrm{m}. The robot starts at position 𝝃0=[−2.5,−1.0]T\boldsymbol{\xi}_{0}=[-2.5,-1.0]^{T}, and the attractor is set to 𝝃a=[2.5,1.0]T\boldsymbol{\xi}^{a}=[2.5,1.0]^{T}.

V-C Obstacle Aware Passivity Using a Robot Arm

A Franka Panda tracks a desired motion around a 0.16 m box, with a 0.12 m margin and a 500 Hz controller. Optical tracking supplies obstacle position. Ten repeated pushed trials per method compare clearance and peak force with an undisturbed reference.

The obstacle-aware trials report about 0.15 m average minimum clearance, compared with below 0.05 m for the velocity-following baseline and several baseline collisions. Improved rejection uses larger momentary forces; Table I preserves their means and standard deviations. Orientation is controlled separately from the translational damping law.

Refer to caption
Refer to caption
((a)) Obstacle-aware controller rejects repulsion and avoids collision
Refer to caption
Refer to caption
((b)) Velocity preserving controller leads to collision with obstacle

Refer to caption

((c)) The two control methods compared with the undisturbed trajectory. The wider line indicates a higher x-value. The darker arrow is the actual, and the desired velocity is the brighter arrow.
Refer to caption
((d)) The specific trajectory is represented by a full line, while the average (dashed line) and variance (shaded area) are evaluated over ten epochs. The mean and variance of the control force are evaluated in the logarithmic space.
Fig. 11: The robot arm, guided by the obstacle-aware passive controller, effectively avoids the disturbance towards the obstacle while maintaining a margin of 0.16 m0.16\text{\,}\mathrm{m} around the obstacle. The experiment was repeated ten times with a similar disturbance applied to the robot arm in each run.
Equation 36
𝝉q=𝑱†​(𝒒)​[𝒟​(𝝃)​(𝒇​(𝝃)−𝝃˙)pα​(𝜶−𝜶a)]\boldsymbol{\tau}_{q}=\boldsymbol{J}^{{\dagger}}(\boldsymbol{q})\begin{bmatrix}\mathcal{D}(\boldsymbol{\xi})\left(\boldsymbol{f}(\boldsymbol{\xi})-\dot{\boldsymbol{\xi}}\right)\\ p^{\alpha}(\boldsymbol{\alpha}-\boldsymbol{\alpha}^{a})\end{bmatrix} (36)
Obstacle Velocity Undisturbed
Closest Distance [mm] 105 ±\pm 21 18 ±\pm 21 167 ±\pm 9
Maximum Force [N] 4.12 ±\pm 0.25 3.58 ±\pm 0.24 2.86 ±\pm 0.37
TABLE I: The mean and standard deviation of the closest distance and maximum force are evaluated over ten epochs for the proposed obstacle aware controller, and the baseline velocity following controller [10]. The undisturbed trajectory is provided as reference values. During the experiments, the obstacle-aware damping avoids collision with an increased minimal distance from the obstacle and, hence, a safer trajectory. This is achieved through momentarily higher forces.

VI Discussion

The experiments support selective disturbance rejection near known obstacles. Stability, tracking accuracy, compliance, and collision avoidance are separate properties with different assumptions and tradeoffs.

VI-A Discrete Control Convergence

The controller’s boundedness result does not ensure convergence to a particular position. Adding a proportional term can promote attraction but changes the analyzed system and requires a new stability argument. The desired velocity field is intended to supply convergence behavior in practice.

Equation 37
𝝉𝒄=𝒈⁡(𝝃)+𝒟⁡(𝝃)​(𝒇⁡(𝝃)−𝝃˙)+𝒦⁡(𝝃)​(𝝃−𝝃a)\boldsymbol{\tau_{c}}=\boldsymbol{g}(\boldsymbol{\xi})+\mathcal{D}(\boldsymbol{\xi})\left(\boldsymbol{f}(\boldsymbol{\xi})-\boldsymbol{\dot{\xi}}\right)+\mathcal{K}(\boldsymbol{\xi})\left(\boldsymbol{\xi}-\boldsymbol{\xi}^{a}\right) (37)

VI-B Applicability and Theoretical Analysis

The analysis motivates tank-free damping control with bounded desired velocities, subject to the stated model and matrix assumptions. It does not automatically extend to arbitrary adaptive stiffness or extra feedback terms.

VI-C Point Mass Agents

The analysis treats a point representation of the robot. Evaluating the controller at several points along an arm is suggested for fuller body coverage, but is not a demonstrated whole-body collision guarantee.

VI-D Caution in Obstacle’s Proximity

Obstacle location is assumed known accurately. With uncertain perception, increasing resistance near an estimated surface may conflict with gentle exploration. Combining avoidance with cautious contact is an open direction.

-E Discrete Time Behavior

The appendix studies a forward-Euler model. Larger damping improves rapid correction but can create oscillation or divergence when the time step is too large. The paper states the bound sᵢ ≤ 2m/Δt, with m the smallest inertia eigenvalue, and illustrates different choices.

Reading note: in the scalar constant case, the velocity multiplier is 1 − Δt·s/m. At equality it is −1, which is marginal rather than strictly decaying. Strictly inside the bound gives decay for that fixed linear case; the boundary and general state-dependent implementations require care. These sampled-data observations should be kept distinct from the continuous-time analysis.

Original figure
Fig. 12: An agent has the desired velocity 𝒇⁡(𝝃)=[1,0]T\boldsymbol{f}(\boldsymbol{\xi})=[1,0]^{T} (gray arrow), and a control matrix 𝒟\mathcal{D} with damping values equal in all directions and the smallest eigenvalue of the inertia matrix mm. The agent is disturbed (purple arrow) position 𝝃0=[0,0]T\boldsymbol{\xi}_{0}=[0,0]^{T} and has a velocity of 𝝃1=[1,1]T\boldsymbol{\xi}_{1}=[1,1]^{T} after the impact.
High damping leads to unstable behavior with increasingly high oscillations. Conversely, the lowest value leads to more deviation from the initial straight line resulting from the higher impedance. The critical value of si=2.0​m/Δ​ts_{i}=2.0m/\Delta t results in stable behavior with immediate correction to the desired velocity.
Equations 38–42
[𝝃t+1𝝃˙t+1]=[𝝃t𝝃˙t]+Δ​t​[𝝃˙t𝝃¨t]\begin{bmatrix}\boldsymbol{\xi}_{t+1}\\ \dot{\boldsymbol{\xi}}_{t+1}\end{bmatrix}=\begin{bmatrix}\boldsymbol{\xi}_{t}\\ \dot{\boldsymbol{\xi}}_{t}\end{bmatrix}+\Delta t\begin{bmatrix}\dot{\boldsymbol{\xi}}_{t}\\ \ddot{\boldsymbol{\xi}}_{t}\end{bmatrix} (38)
[𝝃t+1𝝃˙t+1]=[𝝃t+Δ​t​𝝃˙t𝝃˙t+Δ​t​ℳ−1​(𝒟⁡(𝒇⁡(𝝃t)−𝝃˙t)−𝒞)]=[𝑰𝑰​Δ​t𝟎𝑰−Δ​t​ℳ−1​𝒟]​[𝝃t𝝃˙t]+[𝟎Δ​t​ℳ−1​𝒟]​𝒇^​(𝝃t)\begin{split}&\begin{bmatrix}\boldsymbol{\xi}_{t+1}\\ \dot{\boldsymbol{\xi}}_{t+1}\end{bmatrix}=\begin{bmatrix}\boldsymbol{\xi}_{t}+\Delta t\;\dot{\boldsymbol{\xi}}_{t}\\ \ \dot{\boldsymbol{\xi}}_{t}+\Delta t\,\mathcal{M}^{-1}\left(\mathcal{D}\left(\boldsymbol{f}(\boldsymbol{\xi}_{t})-\dot{\boldsymbol{\xi}}_{t}\right)-\mathcal{C}\right)\end{bmatrix}\\ &=\begin{bmatrix}\boldsymbol{I}&\boldsymbol{I}\Delta t\\ \boldsymbol{0}&\boldsymbol{I}-\Delta t\mathcal{M}^{-1}\mathcal{D}\end{bmatrix}\begin{bmatrix}\boldsymbol{\xi}_{t}\\ \dot{\boldsymbol{\xi}}_{t}\end{bmatrix}+\begin{bmatrix}\boldsymbol{0}\\ \Delta t\,\mathcal{M}^{-1}\mathcal{D}\end{bmatrix}\hat{\boldsymbol{f}}(\boldsymbol{\xi}_{t})\end{split} (39)
𝝀1=eig​(𝑰)𝝀2=eig​(𝑰−Δ​t​ℳ−1​𝒟)\boldsymbol{\lambda}_{1}=\text{eig}(\boldsymbol{I})\qquad\boldsymbol{\lambda}_{2}=\text{eig}\left(\boldsymbol{I}-\Delta t\,\mathcal{M}^{-1}\mathcal{D}\right) (40)
ℳ>0,𝒟>0⇒𝝀2,i>0i=1…N\mathcal{M}>0\;,\;\;\mathcal{D}>0\qquad\Rightarrow\qquad\boldsymbol{\lambda}_{2,i}>0\quad i=1...N (41)
max⁡(𝝀2)≤1⇒si≤2​min⁡(eig​(ℳ))/Δ​t\max\left(\boldsymbol{\lambda}_{2}\right)\leq 1\quad\Rightarrow\quad s_{i}\leq 2\min\Bigl(\text{eig}\bigl(\mathcal{M}\bigr)\Bigr)/\Delta t (42)
Original figure
Fig. 13: An agent with the desired dynamics of 𝒇⁡(𝝃)=𝑹⁡(π/6)​(𝝃−𝝃a)\boldsymbol{f}(\boldsymbol{\xi})=\boldsymbol{R}(\pi/6)(\boldsymbol{\xi}-\boldsymbol{\xi}^{a}) where 𝑹⁡(⋅)∈ℝN×N\boldsymbol{R}(\cdot)\in\mathbb{R}^{N\times N} is the rotation matrix, and mm is the mass of the agent. We assume equal damping values sis_{i}. The controller with a critical stiffness of 2.0​m/Δ​t2.0m/\Delta t leads to fast convergence and a stable system. With lower damping (top left), there is a large drift of the system, which converges to a limit cycle. The high damping of 2.3​m/Δ​t2.3m/\Delta t leads to an unstable system. Interestingly, with damping of 2.1​m/Δ​t2.1m/\Delta t in combination with the nonlinear dynamics, we obtain a visibly stable system.
Concept demonstration

Open image ↗