Literature
Display
← All papers

BiFlex: A Passive Bimodal Stiffness Flexible Wrist for Manipulation in Unstructured Environments

Gu-Cheol Jeong · Stefano Dalla Gasperina · Ashish D. Deshpande · Lillian Chin · Roberto Martín-Martín

· IEEE Robotics and Automation Letters ↗

Robotics5 min text · CondensedOriginal paper ↗
Contents

At a glance

BiFlex is a passive wrist whose printed honeycomb structure stays relatively stiff below a designed buckling load and becomes compliant above it. A universal joint constrains rotation to two bending directions. Three versions support different grippers, targeting less than 1 cm fingertip deflection for objects up to 500 g. Bench and manipulation tests demonstrate a useful precision–compliance tradeoff, with no added wrist actuators or sensing needed to switch stiffness.

I Introduction

Uncertain object and surface geometry can make position-controlled rigid arms generate large contact forces. Constant compliance helps contact but sacrifices placement precision; actively variable stiffness adds hardware and control complexity. BiFlex instead sets the transition mechanically: ordinary manipulation stays below the buckling threshold, while stronger contact induces passive bending. The contribution is a tunable structure, approximate design analysis, and tests across three grippers—not a general certification of safe interaction.

Refer to caption
Refer to caption
Fig. 1: (top) The BiFlex wrist, featuring a honeycomb structure, mounted on a robotic arm and attached to a robotic gripper. (bottom) Typical manipulation sequence using the BiFlex wrist. When grasping in unstructured environments, the wrist transitions from high- to low-stiffness states upon contact with the table surface to exploit it as a constraint. Illustration adapted from [5].

II Related Work

Actuated wrists use motors, tendons, or pneumatics to change stiffness continuously. Underactuated designs can switch between two modes but still need powered locking or switching. BiFlex removes that active mechanism. Its tradeoff is that stiffness changes with the applied load and manufactured geometry, rather than being independently commanded during a task.

Refer to captionRefer to captionDeflectionForce / TorquePrecision ModeCompliant ModeBuckling Point
Fig. 2: The desired bimodal behavior ensures high stiffness during precision mode (red) and low stiffness during compliant mode (green), with a smooth transition at the buckling point (yellow).

III BiFlex Wrist Design

The wrist combines a buckling honeycomb, a central universal joint, and customizable mounting plates. The intended behavior is small elastic deformation during precision motion, larger bending under contact, and recovery after unloading. The universal joint suppresses axial rotation while allowing two bending degrees of freedom.

Refer to caption
Fig. 3: (left) Structural configuration of the BiFlex wrist design. The top plate and base frame are customizable to accommodate different robotic arms and grippers. (right) Cross-sectional and frontal views of the honeycomb structure.

III-A Buckling Honeycomb Structure

Twelve modules are distributed around the wrist. The design is at most 40 mm tall including adapters, with a roughly 21 mm flexible module and up to 40° bending. A honeycomb offers more tuning variables than the initially tested single-beam structure. The diagonal beam angle and width control both pre-buckling stiffness and the load at which the compliant regime begins. Rotational symmetry is a design aim; the central joint also helps constrain unwanted yaw.

III-B Analysis of the Honeycomb Buckling Behavior

The analysis treats one module under vertical load, assumes symmetric force sharing, neglects the force perpendicular to diagonal beams, and models beams using axial stiffness and a critical buckling load. It derives equivalent stiffness from diagonal and central beam deformation, then assumes the diagonal beams buckle first and trigger module buckling. The preserved equations show how geometry and Young’s modulus determine the transition. Beam width and diagonal angle are the principal tuning variables. These simplified relations guide fabrication; they are not an exact model of the full printed three-dimensional structure.

Refer to caption
Fig. 4: Frontal view (left) and free-body diagram (right) of one buckling honeycomb module. 𝐅𝐢\mathbf{F_{i}} represents the compression force applied to each single honeycomb module.
Equations 1–5
Fi=Ke​q​δyF_{i}=K_{eq}\delta_{y} (1)
δx=Fa​sin⁡γK2=Fi​sin⁡γ2​K2​cos⁡γ\delta_{x}=\frac{F_{a}\sin\gamma}{K_{2}}=\frac{F_{i}\sin\gamma}{2K_{2}\cos\gamma} (2)
Fi2​cos⁡γ=K1⋅(δy​cos⁡γ−δx​sin⁡γ)\frac{F_{i}}{2\cos\gamma}=K_{1}\cdot(\delta_{y}\cos\gamma-\delta_{x}\sin\gamma)
δy=1cos⁡γ​(Fi2​K1​cos⁡γ+Fi​sin⁡γ2​K2​cos⁡γ)\delta_{y}=\frac{1}{\cos\gamma}\left(\frac{F_{i}}{2K_{1}\cos\gamma}+\frac{F_{i}\sin\gamma}{2K_{2}\cos\gamma}\right)
Ke​q=K2+K1​sin⁡γ2​K1​K2​cos2⁡γK_{eq}=\frac{K_{2}+K_{1}\sin\gamma}{2K_{1}K_{2}\cos^{2}\gamma} (3)
K1\displaystyle K_{1} =E​Ah=E​b​Lh=2​E​b​L​c​o​s​γH\displaystyle=\frac{EA}{h}=\frac{EbL}{h}=\frac{2EbLcos\gamma}{H}
K2\displaystyle K_{2} =E​Achc=E​bc​Lhc\displaystyle=\frac{EA_{c}}{h_{c}}=\frac{Eb_{c}L}{h_{c}}
Ke​q=H​bc+2​b​hc​cos⁡γ​sin⁡γ2​E​L​b​bc​cos3⁡γ\boxed{K_{eq}=\frac{Hb_{c}+2bh_{c}\cos\gamma\sin\gamma}{2ELbb_{c}\cos^{3}\gamma}} (4)
Pc​r=π2​E​Ih2=π2​E​b3​L12​h2P_{cr}=\frac{\pi^{2}EI}{h^{2}}=\frac{\pi^{2}Eb^{3}L}{12h^{2}}
Fi≥2​π2​E​b3​L​cos⁡γ12​h2\boxed{F_{i}\geq\frac{2\pi^{2}Eb^{3}L\cos\gamma}{12h^{2}}} (5)

IV Instantiation to Different Robotic Grippers

Versions for the Franka Hand, Robotiq 2F-85, and BaRiFlex account for different gripper masses and lever arms. The target is a 500 ± 50 g supported object with less than 1 cm fingertip displacement; acceptable buckling torque is within 10% of the design value. TPU-95A forms the flexible honeycomb, and PLA forms the customized plates. Total printing takes about 10 hours, including four for the honeycomb. Table I retains dimensions and selected beam parameters for each gripper.

TABLE I: Gripper characteristics and design parameters.
Gripper dimensions Buckling point BiFlex parameters
Weight
[kg\mathrm{k}\mathrm{g}]
Length
[mm\mathrm{m}\mathrm{m}]
Torque
[Nm\mathrm{N}\mathrm{m}]
Angle
[°]
Width
(bb) [mm\mathrm{m}\mathrm{m}]
Angle
(γ\gamma) [°]
Franka 0.70 135 0.96 4.20 0.90 50
Robotiq 1.10 155 1.325 3.70 1.00 20
BariFlex 0.75 205 1.378 2.80 1.20 5
Refer to caption
Refer to caption
Refer to caption
Refer to caption
(a) Franka Hand
Refer to caption
(b) Robotiq 2F-85
Refer to caption
(c) BariFlex [26]
Fig. 5: Honeycomb structure designs corresponding to the three different grippers. Each design is tailored to the specific gripper, with the angle γ\gamma and diagonal beam width bb adjusted to achieve the desired buckling point.

V Characterization

A compression machine loads a cantilever attached to each wrist; force and displacement are converted to torque and angle. Three tests per wrist are averaged. Curves rise approximately linearly before leveling after buckling. Measured angle/torque pairs are 3.06°/1.51 Nm for BaRiFlex, 3.99°/1.45 Nm for Robotiq, and 3.40°/0.95 Nm for Franka, within the stated design bounds. The buckling point is an operational estimate from an 80%-torque interpolation line and the plateau, rather than a directly observed binary switch.

Refer to caption
Original figure
Fig. 6: Compression test. (left) The compression testing machine applies a downward force on a cantilever beam connected to the BiFlex, deforming it. (right) Results indicate that the buckling point (∙\bullet) falls within its tolerance area (hatch-line box) around the desired buckling point (★\bigstar).
Refer to caption
Original figure
Fig. 7: Pressing test. (top) The robot, equipped with an F/T sensor, BiFlex, and different robot grippers moves downwards, contacting an aluminum frame with the finger tips and causing BiFlex to deflect. (bottom) Loading and unloading curves with measured torque as a function of the deflection, including experimental buckling point (∙\bullet), the designed buckling point (★\bigstar), and the design tolerance area (hatch-line box).

VI Experimental Evaluation

Experiments separately test contact-load mitigation, surface following, placement accuracy, and grasping with a table constraint. These are bounded demonstrations using a Franka arm and selected grippers, not exhaustive tests of impact, fatigue, or human contact.

Refer to captionSpongeGripperBiFlexRobot ArmTilted Surface30°30^{\degree}+y+x

Refer to caption
Fig. 8: Wiping test. (left) The robot, equipped with an F/T sensor and a Robotiq hand holding a sponge, moves horizontally with a position controller to wipe a triangular surface, starting at different heights. We compare the behavior with and without BiFlex. Shape profile (top), measured forces (middle), with a 15 N15\text{\,}\mathrm{N} force safety limit (red dotted line), and measured torques (bottom) during the wiping motion for the case with BiFlex (left column) and without (right column). The robot equipped with BiFlex successfully wipes the triangular surface up to a height of 50 mm, whereas without BiFlex, the maximum height before reaching the safety force threshold is 14 mm.

VI-A Mitigation of Contact Wrenches: Pressing Test

Each gripper presses against a frame until the wrist bends 10°, then unloads; a six-axis sensor measures transmitted forces and torques over three trials. Franka and Robotiq buckling points fall within the 10% design tolerance. BaRiFlex deviates, plausibly because the soft hand adds deformation that the wrist-only model omits. The authors’ 15 N threshold is a force threshold, although one sentence incorrectly calls it a torque margin. Results support load reduction under this pressing setup.

VI-B Adaptation in Contact-Rich Interaction: Wiping Test

With the same stiff position controller, a Robotiq hand wipes 30° triangular slopes using a 5 mm sponge. BiFlex traverses heights up to 50 mm while staying below the experiment’s 15 N force cutoff; a rigid connection reaches only 14 mm. Buckling starts for triangles of at least 20 mm, while lower profiles use the stiffer regime. This demonstrates useful passive adaptation without force-feedback control for these surfaces.

VI-C Maintaining Accuracy: Pick-and-Place Test

Fifteen household objects are tested with three placement attempts each. Fourteen objects below 500 g are placed without the defined shelf-contact failure. A 600 g water bottle triggers buckling and more than 10 mm fingertip displacement, causing shelf contact. The test therefore illustrates the designed load boundary as well as successful placement below it; it does not establish precision for arbitrary payloads, speeds, or orientations.

Refer to caption
Fig. 9: Pick-and-place test. Objects used during our experiments (top-left) and sequence of steps of the pick and place execution (top-right). (bottom) A successful demonstration of precision stage with object <<500g (bottom-left), whereas a 600g object induced buckling with over 10mm fingertip deflection (bottom-right). For all objects under 500g, the robot can accurately place them in the shelf without collision, as per our design.

VI-D Grasping with Environmental Constraints

A hex wrench is grasped by sliding fingers against the table. Commanded grasp heights range from the true height to 50 mm below the surface in 5 mm steps. BiFlex succeeds throughout the tested range and recovers after lifting. The rigid wrist succeeds only at the true height and 5 mm below before exceeding the force limit. The table supplies the geometric constraint, while wrist compliance absorbs the commanded height error.

Refer to caption
Fig. 10: Grasping with environmental constraints. (top) Pictures show task phases and BiFlex modes: high- to low-stiffness upon table contact, then back to high stiffness when lifting the object. (bottom) Force and torque measurements for the largest simulated perceptual error: the robot assumes the object is 5 cm under the table. BiFlex compensates for this error, using the environment as a constraint while keeping forces and torques within safety margins.

VII Discussion and Conclusion

Passive buckling can combine useful free-space precision with contact compliance in a compact replaceable wrist. Remaining limitations include low-frequency oscillations during wiping, deviations from universal-joint tolerances, and material wear that may change stiffness over time. Replacing the printed honeycomb is practical but does not eliminate those effects. Broader safety, durability, and dynamic-load claims require further evaluation.

Concept demonstration

Open image ↗