Organizations: Instituto de Investigación en Ingeniería de Aragón (I3A), University of Zaragoza, Spain · Université Clermont Auvergne, Clermont Auvergne INP, CNRS, Institut Pascal, F-63000 Clermont-Ferrand, France
Industrial manipulators offer reach and precision but can damage nearby structures in confined spaces. Soft robots provide compliant contact but typically have limited reach. Rigid-soft systems combine their advantages, although they often require dedicated actuation and control hardware. We propose actuating a passive tendon-driven soft extension through manipulator joint motion. As the robot moves, the exterior spans of tendons anchored to upstream links change, altering tendon tension and deforming the extension. Inspection through narrow apertures, for example, constrains the motion of the end effector (EE). We develop a controller that regulates the soft tip while preserving such EE constraints. A differential model separates rigid transport from soft deformation, combining analytic arm and tendon kinematics with a locally identified span-to-deformation map. We also derive a local upper bound on the number of independent deformation directions transmitted by the tendons under the EE constraint, and conditions for the exponential convergence of the tip error. A quadratic program with control barrier functions imposes obstacle clearance, fulcrum limits, and model-validity conditions. Simulations and real experiments show tip and shape regulation under free, position-constrained, and fulcrum-constrained EE motion.
Figures & tables
Fig. 1: Three robot extension approaches for confined access: a passive rigid extension, an active soft extension, and the passive soft extension proposed here. The active design requires dedicated actuation, power, and control hardware, whereas the proposed design uses manipulator joint motion to actuate its tendons.
Fig. 2: Manipulator-driven actuation of a passive soft extension in a UR10 simulation. Top: four states under a laparoscopic fulcrum constraint. Two tendons run from anchors on the forearm and wrist (circles) to entry points at the base of the soft extension. The tip is black, and its path is blue. Bottom: corresponding joint angles, exterior tendon spans, and tip position. The spans change by up to 18mm and the tip moves by up to 170mm .
Approach
Actuator-free
Shape regulation
Industrial manipulator interface
Actuated by
Continuum / tendon-driven [ 1 , 4 ]
no
yes
no
own actuators
Hybrid rigid–continuum [ 7 , 8 ]
no
yes
yes
own actuators
Passive adaptive tools [ 9 , 10 ]
yes
no ∗
yes
contact / arm motion
Deformable-object shape control [ 11 , 12 ]
– †
yes
yes
grasp motion
Passive soft extension (this work)
yes
yes
yes
arm joint motion
TABLE I: Actuation and feedback characteristics of related approaches and of the passive soft extension.
Fig. 3: Notation. Left: the Jacobian JD (gray box) maps joint rates to the motion of the distal frame through two parallel paths whose contributions add, as in ( 9 ). One path is the transport Jtransport of the soft extension with the EE at fixed shape. The other, RˉBJspanJℓq , runs through the spans, tensions, and relative twist. Only its factor Jspan (shaded) depends on the mechanics of the soft extension. Selecting the translational rows of JD gives the tip-position Jacobian Jt=SJD . Right: the simulated UR10 at t=10s in Fig. 2 . Joint motion q˙ moves the anchors ai relative to the entry points bi on the EE frame {B} and changes the exterior spans ℓio along the unit vectors di . The inset magnifies the cross-section of the soft extension at its base (ring), with the entry points at Bbi in {B} . The tendons continue inside the soft extension to its tip (drawn apart for visibility), and their tensions τ deform the soft extension, whose shape z determines the distal frame {D} and the body points pj . The tip velocity is the translational part of the sum of transport, Jtransportq˙ , and deformation, RˉBBVD/B . Frames: x red, y green, z blue.
Fig. 4: Overview of the method. Model (top): the arm kinematics and the measured tendon anchors give the analytic terms Jℓq ( 4 ) and Jtransport ( 8 ), and the span-to-deformation map Jspan and its span domain L are identified from probing motions and corrected online. Together they form the identified model JD ( 10 ). Controller (bottom): in the null-space control law ( 15 ), the EE task (primary) computes the restoring velocity q˙c and the projector Nc from the error ec about its reference, pB⋆ or c (Table II ). The tip task (secondary) uses both to compute the null-space motion Ncν ( 18 ). When barriers are defined, the weighted quadratic program ( 19 ) can replace q˙nom to satisfy the barrier conditions. Plant (right): the robot provides q , and a camera measures the tip pD and the body points pj . In this example, the EE constraint is a fulcrum (purple) through the point c of an aperture, and the grid schematically indicates the tip positions reachable with ec=0 and spans inside L .
Fig. 5: The three EE constraints on the simulated UR10. Each panel compares a reference EE pose (solid) with another pose that satisfies the constraint (faded, over a translucent arm). Free EE: the task constrains no joint motion. Position-constrained EE: pB is fixed and only the orientation can change. Laparoscopic fulcrum: the tool axis u (the z axis of the EE) passes through the point c , the cone (dashed outline) shows the tilt limit of the cone barrier (Table III ), and the arm can pivot, roll, and insert. Table II gives the errors. Frames: x red, y green, z blue.
Constraint
Error ec(q)
rc
Free EE
none
0
Position-constrained EE
pB⋆−pB
3
Laparoscopic fulcrum
(I−uuT)(c−pB)
2
TABLE II: EE constraints (Fig. 5 ).
Fig. 6: The barrier functions of Table III ; dashed lines mark the margins h=εh . Left: obstacle clearance between the surface of the soft extension (body points pj , radius r ) and the obstacle O . Middle: fulcrum barriers in a section through the aperture (center c , radius rap ), for the aperture crossing of the tool axis u ( hap ), its tilt below the cone half-angle θˉ ( hcone ), and the insertion depth δ between δmin and δmax ( hins ; EE dashed at both limits). Right: the spans inside the identified span domain L in simulation, with the spans of one run.
Barrier
Function h
Gradient ∇hT
Obstacle clearance
hobs=dist(P,O)−r
ηTJp
Fulcrum: aperture
hap=(rap−r)cosθ−dc
from JB
Fulcrum: cone
hcone=θˉ−θ
from JB
Fulcrum: insertion
hins=min(δ−δmin,δmax−δ)
from JB
Span domain L
hspan=mini{ℓio−ℓi−,ℓi+−ℓio}
±[Jℓq]i
TABLE III: Barrier functions, h≥εh (Fig. 6 ).
Fig. 7: Convergence of the nominal controller ( 25 ) on the simulated plant for five targets, reached in sequence under the fulcrum, with the damped inverse ( kt=1.4 , ξ=0.04m ), W=I , and no barriers. Along the run, A keeps full row rank and β stays between 0.88 and 0.92 . Top: tip path through the numbered targets, over the soft extension at each target switch, and a legend of the plots. Middle: V on a logarithmic vertical axis, with the exact Jacobian (solid) and with the identified model ( 10 ) (dashed). The bound ( 31 ) (dotted) restarts at each target switch from the current value of V and uses the smallest β of the run. Bottom: the descent rate −V˙/V with both Jacobians and the band [2ktβ,2kt] (shaded) that ( 28 ) predicts for the damped inverse.
Fig. 8: The three simulated scenarios: (a) position-constrained EE, (b) laparoscopic fulcrum, and (c) infeasible targets. Top: snapshots at the start and end of two target intervals. Middle: 3-D view with the targets, the obstacles, the tip path, and the soft extension at the end of each target (thin) and of the run (thick). Bottom: norm of the tip error with the 4mm arrival tolerance, joint rates, and changes of the two exterior spans; dotted lines mark target changes. In (c), T3 lies inside an obstacle and T5 beyond the insertion limit, and the error settles at a nonzero value while the corresponding barrier is active.
Plant
UR10 with a 320mm Cosserat rod [ 40 ] and two tendons, 100Hz
Sensing
Joint states; tip pose and five body points at 20Hz , Gaussian noise of 0.5mm standard deviation
Identification
Jℓq , Jtransport exact; Jspan fitted on an 11×11 grid of spans ( ±10mm ), refined online
TABLE V: Simulated scenarios. All impose the obstacle and span-domain barriers, and the fulcrum scenarios also impose the cone and insertion barriers.
Fig. 9: Experimental platform. Top left: the UR10 on a stationary mobile base, the passive soft extension, and the RGB-D camera, calibrated in the robot base frame with the ArUco markers on the table. Top right: passive soft extensions with parallel tendons (Exps. 1–5) and with helical tendons of one turn each (Exp. 6); colored balls mark three body points and the tip. Bottom: tendon routing from anchors on the upper arm (tendon 1) and the forearm (tendon 2) to the base of the soft extension at the gripper.
Fig. 10: Experiments with (a) a free EE and (b, c) a position-constrained EE. Top: camera snapshots with the detected balls highlighted. Middle: 3-D view with the targets, the obstacles, the tip path, and the soft extension at the end of each target (thin) and of the run (thick). Bottom: norm of the tip error with the 10mm threshold, joint rates, and changes of the two exterior spans; dotted lines mark target changes. The time plots have the same scales as in Fig. 12 .
Fig. 11: EE constraint and barrier values (Table III ) measured in the six experiments. Shading marks negative barrier values, dotted lines the margins εh , and thick bands the intervals in which the barrier condition was active in the QP. (a) Distance of the EE from its reference for the position-constrained EE. (b)–(d) Aperture, cone, and insertion barriers of the fulcrum. (e) Clearance of the tracked body to the obstacles, clipped at 125mm , with the 10mm margin below which the obstacle condition is imposed. (f) Smallest distance of the spans to the bounds of the identified span domain.
Fig. 12: Experiments with the fulcrum at a finite aperture, with parallel tendons (a, b) and with helical tendons (c). Each column shows snapshots, a 3-D view, the tip error, the joint rates, and the span changes in the layout of Fig. 10 . In (b), T3 is infeasible by design.
Plant
UR10; passive foam tube, 475mm long, less than US$10 each; two tendons, anchored at mid upper arm and mid forearm
Sensing
Joint states; RealSense D415 at 30Hz , three body markers and a tip marker
Identification
Linear Jspan from 13 probe poses ( 10 for fitting, 3 for validation), refined online
Control
Joint velocity commands at 125Hz , 8 – 12∘/s at most; kt=1.2 , ξ=0.02m , μ=0.04m ; tip speed ≤20mm/s
Targets
Reached when the steady-state tip error is below 10mm
TABLE VI: Experimental setup.
Exp.
EE constraint
Tendons
Obst.
Targets
1
free
parallel
1
3+2
2
position-constrained
parallel
1
2+1
3
position-constrained
parallel
2
2
4
laparoscopic fulcrum
parallel
1
2
5
laparoscopic fulcrum
parallel
1
3+1
6
laparoscopic fulcrum
helical
0
4
TABLE VII: Experiments. All impose the span-domain barrier and, when obstacles are present, the obstacle barrier. Under the fulcrum, they also impose the three fulcrum barriers (aperture, cone, and insertion). Targets: number predicted feasible + number infeasible by design.
Limiting factor
Exp.–target (error in mm)
Infeasible by design
1–T2 (18), 1–T4 (12), 2–T2 (43), 5–T3 (24)
Span domain
1–T3 (23), 1–T5 (12), 2–T3 (19)
Aperture and insertion
4–T2 (28), 5–T2 (37), 6–T1 (17)
Stationary point (no active barrier)
1–T1 (14), 2–T1 (37), 3–T1 (42), 3–T2 (42)
TABLE VIII: Unreached targets, grouped by limiting factor, with the steady-state tip error. The barriers are those of Table III . Stationary points are possible since the output has 12 coordinates, whereas the rank of A is at most six (Section VII-D ).
Hong Kong Embodied AI Lab, Hong Kong SAR, China · The Chinese University of Hong Kong, Hong Kong SAR, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China