Embodiment Is Doing the Sensing—and the Reweighting—in Two Small-Robot Designs
A whiskered drone and a water-shifting undulatory robot point to a shared design strategy: move capability into the body, then keep onboard decisions tightly tied to what that body can reveal.
By Felix Park · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human engineering credentials or firsthand experience.
Key points
- Delft researchers report that artificial whiskers let a tiny drone estimate contact depth and location, navigate without vision, and run its tactile-processing pipeline onboard with limited memory.
Sources: S1
- NYU Tandon’s WorMa changes the distribution of water between its head and tail to alter traction and body posture; its successful strategy changes with the obstacle.
Sources: S2
- The comparison suggests a practical rule for embodied robotics: a simpler controller can be valuable only when the body exposes useful, timely signals and can physically make the adjustment the controller selects.
Two versions of complexity moved into the body
Small robots are often presented as a choice between adding sensors and accepting less autonomy. The two developments here offer a different proposition. A flying robot from Delft University of Technology uses flexible, contact-sensitive whiskers to obtain local information where cameras and ranging sensors can be poorly suited to darkness, dust, smoke, weight limits, and onboard resource limits. NYU Tandon’s WorMa, meanwhile, changes where water sits inside its undulating body, using that mass shift to change contact force, traction, and the sequence by which it negotiates terrain. Neither approach eliminates control software. Each instead gives the controller a body whose mechanics produce a useful signal or useful change in state.
What the drone can actually observe
The drone’s whiskers are an explicitly local sensing system. A pair mounted at the front contacts a surface, and pressure sensors at each whisker base register bending. The reported processing uses those changes to estimate the depth and location of contact in space. That supports obstacle avoidance, following a surface, and tactile mapping. In the reported experiments, the vehicle followed rigid and soft contours in complete darkness, explored enclosed spaces by wall following, and found an exit while building tactile maps without cameras or other vision-based sensors.
Sources: S1
The resource constraint is central to the claim. The source describes drones below a stated mass threshold as constrained in sensing, computing power, and energy, while conventional cameras, LiDAR, and rangefinders may be too heavy or unreliable in poor visibility. The Delft system filters propeller-induced airflow disturbance, drift, and distortion from whisker signals, runs on a microcontroller, and uses a stated amount of memory. The important accomplishment is therefore not simply that the drone touches a wall. It is that it tries to distinguish environmental contact from the turbulence produced by its own act of flying, without offboard computing or positioning.
Sources: S1
Inference: this is embodied perception rather than passive collision detection. The whisker’s bend turns an otherwise ambiguous close encounter into a structured measurement, but only after onboard processing separates contact from self-generated aerodynamic noise. Its useful world model is correspondingly narrow: it is rich near the whiskers and during contact, rather than a broad visual view of an entire room. That is a sensible trade when the mission is to explore an unknown enclosed route under low visibility, but it does not establish performance in every aerial navigation setting.
Sources: S1
Sources: S1
What WorMa changes when it cannot rely on one balance
WorMa’s key variable is not a sensory whisker but internal mass placement. The robot pumps water between head and tail while retaining the same basic gait. On an incline, moving water toward the head increases the force at front contact points and improves traction. The supplied account says that head-biased water placement alone carried the robot up the steepest tested incline, while other placements slipped backward; it also reports a lower cost of transport than the other configurations.
Sources: S2
A step exposes why a single body configuration is insufficient. According to the report, neither a permanently head-heavy nor a permanently tail-heavy version cleared a step by itself. WorMa instead uses a sequence: water begins at the head for traction, moves to the tail as the head and neck lift and the tail pushes the robot nearer, then returns to the head after the head anchors on the edge. The reported result is step climbing at the stated tested height. This is not merely ballast adjustment; it is a terrain-specific reallocation of mechanical advantage.
Sources: S2
Inference: WorMa’s controller appears to face a decision problem different from the drone’s. The drone first needs to determine whether a deflection reflects a real surface or its own propeller wash. WorMa needs to select a mass distribution and body phase that make the next contact useful. Both systems are limited by communication inside the robot: the drone must carry a reliable contact signal from whisker base to onboard decision logic, while WorMa must move water from one body region to another before its contact geometry changes. In both cases, delayed, noisy, or incomplete inputs can turn an embodied advantage into a failure mode.
The shared lesson is not “less hardware”
These projects differ sharply in what embodiment is buying. The whiskered drone substitutes a lightweight contact channel for sensing that can fail in low visibility, then converts that channel into local geometry. WorMa uses internal fluid to change the forces that its existing contacts can generate. One design makes the environment more observable; the other makes the robot more mechanically adaptable. Treating them as the same achievement would obscure the practical distinction between sensing a constraint and altering the body to meet it.
Their common lesson is more specific: body design can reduce the demand placed on a general-purpose perception stack or a highly elaborate gait library, but it cannot remove the need for state estimation and timing. The Delft team had to compensate for airflow created by the drone itself. WorMa needed different water placements for slopes and steps, rather than a universal configuration. Embodiment is therefore not a shortcut around uncertainty. It is a way to shape the uncertainty into a form the robot can act on with constrained onboard resources.
What would change this assessment is evidence beyond the reported demonstrations. For the drone, useful next evidence would include results across different surface materials, contact speeds, airflow conditions, and routes where a whisker cannot reach the relevant obstacle. For WorMa, the decisive additions would be performance across varied terrain and fluid-management conditions, along with evidence of how the robot detects which terrain sequence to choose rather than following a preselected one. The supplied material reports successful behaviors and specific tested conditions, but it does not settle those broader operational questions.
Why it matters
The practical design question is not whether robots should be sensor-rich or mechanically clever. It is whether a body-level feature supplies the right observation or force adjustment before limited computation, energy, and internal actuation become the bottleneck. The drone and WorMa show two distinct answers: touch can make nearby space legible when vision is unreliable, while movable internal mass can make a gait viable when fixed weight distribution fails.
Sources
- Bio-inspired whiskers enable tiny drones to navigate in darkness using touch — Tech Xplore Robotics ·
- Robot Videos: Quadrotor From Birotors, Lunabotics, More — IEEE Spectrum Robotics ·