A lightweight water-surface robot driven by inchworm-inspired soft pneumatic joints.
ไธญๆ็ๆฌ่ฏท่ง README.zh-CN.md
Sampling a body of water is an unavoidable step in aquatic environmental research, and a tedious one โ large bodies of water eat enormous amounts of labour. Existing water-surface robots are mostly large rigid platforms with high centres of gravity and poor access to confined spaces.
This is a small, cheap, floating alternative. Four silicone soft joints deform under internal air pressure to drive four limbs; passive louvre-type webbed feet convert that motion into net forward thrust. It floats, it steers, it takes a water sample, and it does all of it over Bluetooth from a phone.
| Hull | 200 ร 130 ร 67 mm, boat-shaped |
| Actuation | 12 V air pumps โ solenoid valves โ silicone soft joints |
| Controller | Lolin32 lite (ESP32) + custom PCB shield |
| Comms | Bluetooth LE, Blinker IoT platform |
| Speed | โ 0.04 m/s in still water |
| Joint deflection | 190ยฐโ195ยฐ (gen-3) vs โ100ยฐ (gen-2) |
| Iterations | 3 soft-joint generations ยท 5 webbed-foot generations |
| Timeline | Nov 2022 โ May 2024 |
Ten prior works were surveyed (full review). They fall into three camps, and each one ruled out a design direction:
| Camp | Example | Why not |
|---|---|---|
| Large rigid surface craft | 7 m autonomous cleaning boat | Size kills manoeuvrability and access |
| Frog jumping robots | Detonation-driven, spring-loaded, geared linkages | High centre of gravity, demanding motor sync, poor on water |
| Smart-material actuators | Dielectric elastomer, water-hydraulic muscle | Need high voltage or high pressure โ bad underwater |
That left a gap: small + surface-swimming + low-pressure pneumatic + fully passive feet. That's what this robot is.
Battery โโฌโ Large pump โโฌโ Accumulator ร2 โโฌโ Valve โ left-rear joint
โ โ โโ Valve โ right-rear joint
โ
โโ Small pump โโฌโ Accumulator ร2 โโฌโ Valve โ left-front joint
โ โโ Valve โ right-front joint
โ โโ Valve โ sampling tube
Every limb attaches through the same elliptical cylindrical interface (15 ร 25.7 ร 20 mm, โ12 mm bore). One part does three jobs: structural mount, air passage, and sealing face. Standardising it is what made rapid iteration possible โ limbs swap onto a test rig in seconds.
๐ Overview ยท Pneumatics ยท Electronics ยท Firmware
The webbed feet went through five generations. The first three failed, and they failed for a reason I didn't understand until the fourth.
| Gen | Design | Outcome |
|---|---|---|
| 1 | Rigid fan-shaped paddle | โ Huge drag on recovery stroke |
| 2 | Fan shape with curved water-facing surface | โ Curvature nowhere near enough |
| 3 | Fully silicone flexible paddle | โ Cut drag โ and cut thrust with it |
| 4 | Two-layer one-way valve (wooden lattice + plastic sheet) | |
| 5 | Trapezoidal louvre, multiple vanes on internal pivots | โ Low opening force, large open area, passive |
The diagnosis that unlocked it came at gen 4:
The problem was never that thrust was too small. Thrust and recovery drag were equal. The foot has to deform during the stroke so that its water-facing area shrinks on the way back โ otherwise every gram of forward push is cancelled on return.
And because an actively-driven foot would have needed its own air supply and control channel, the solution had to be entirely passive. Gen 5 louvres close under forward pressure to form a solid paddle, and blow open on the return stroke to unload the limb. No actuator, no valve, no code.
๐ Full webbed-foot iteration history
Replacing the joint's smooth inner wall with an inchworm-like corrugation roughly doubled deflection.
| Gen 2 โ smooth inner wall | Gen 3 โ corrugated inner wall |
|---|---|
 |
 |
| Sample | Gen 2 deflection | Gen 2 fill time | Gen 3 deflection | Gen 3 fill time |
|---|---|---|---|---|
| 1 | 105ยฐ | 0.88 s | 195ยฐ | 1.65 s |
| 2 | 100ยฐ | 0.79 s | 195ยฐ | 1.73 s |
| 3 | 100ยฐ | 0.84 s | 190ยฐ | 1.47 s |
| Mean | 101.7ยฐ | 0.84 s | 193.3ยฐ | 1.62 s |
Three samples per generation from different casting batches; measured by frame-by-frame video analysis. The trade is real โ you buy 1.9ร the range with 1.9ร the fill time โ and for a low-frequency swimming gait that's a good deal.
๐ Soft joint iteration history ยท Abaqus simulation
- Print the moulds โ
hardware/cad/mold/, four parts: outer frame, centre post, inner ring, end plate - Cast the joints โ Shore-0 silicone; see process photos
- Print structural parts โ
hardware/cad/parts/ - Assemble pneumatics โ BOM and pneumatics guide
- Flash the firmware โ
firmware/frog-main/, see firmware README - Pair over Bluetooth with the Blinker app and map four buttons
Works:
- Stable flotation with a low waterline; feet mostly submerged, giving useful frontal area
- Reliable inflate/deflate cycling on the water
- Directional control (forward / left / right) and water sampling over BLE
- โ 0.04 m/s in still water
Doesn't, and I want to be clear about it:
- Speed is low. 0.04 m/s is a proof of concept, not a useful survey platform.
- Joint sealing is a hack. PTFE tape โ rubber ring โ cable tie, three passive layers. It works far better than the original bare fit, but it is not a designed seal.
- Foot testing was qualitative only. Generations 3โ5 were compared by dragging them through a pool by hand at roughly matched speed. There is no force-gauge data. This is the weakest part of the experimental design and the first thing I'd fix.
- The pumps are heavy. Swapping them for a compressed-gas cylinder was recommended early on and never happened.
๐ Test data
Roughly 18 months, Nov 2022 โ May 2024. Full timeline
- ๐ฅ National Silver Medal โ ICC Global Inventors Challenge, China National Final, 2023
- ๐ Outstanding Student โ China Teenager Science & Technology Talent Program (่ฑๆ่ฎกๅ)
- ๐ Regeneron ISEF 2024 โ selected for Team China through the CAST national screening
docs/ Technical write-ups, figures, bill of materials
firmware/ Arduino/ESP32 source โ frog-main is the final version,
legacy/ preserves the six earlier iterations
hardware/ STL / DXF / SLDPRT models and assembly drawings
simulation/ Abaqus finite-element results for the soft joint
media/ Demo GIFs, hero photos, silicone casting process
paper/ Thesis (Chinese, full text) and bilingual abstracts
See CITATION.cff, or:
Lyu, H. (2024). Bionic Frog Robot: A Lightweight Water-Surface Robot Driven by Inchworm-Inspired Soft Pneumatic Joints. https://github.com/purticra/bionic-frog
| Content | Licence |
|---|---|
firmware/ |
MIT |
hardware/ |
CERN-OHL-S-2.0 |
docs/, media/, simulation/ |
CC BY-NC-SA 4.0 |
paper/ |
CC BY-NC-ND 4.0 |
Advisor Prof. Pan Likun, East China Normal University, whose lab hosted the summer development work and whose early diagnosis โ that airtightness, not actuation, was the binding constraint โ shaped the whole project.
Thanks also to the graduate students at the ECNU Optics Building, and to the China Teenager Science & Technology Talent Program for access to university facilities.