Bones and Skin: Structural Materials, Stress Analysis, and the Art of Building a Body That Survives
DOI: 10.5281/zenodo.18960243 · View on Zenodo (CERN)
Series: Open Humanoid — Article 5 of 20 Author: Oleh Ivchenko Date: March 2026
The Moment Everything Goes Wrong
Picture this: an 80-kilogram humanoid robot loses its footing on a slightly damp factory floor. The fall takes roughly 0.4 seconds — the time it takes you to blink twice. By the time the software stack registers what is happening, the hip joint is already at 40 degrees of unexpected lateral deflection. Then comes the impact.
At that instant, nothing matters less than the motion planning algorithm. The reinforcement-learned gait policy is irrelevant. The 1 kHz control loop cannot help you now. What matters — the only thing that matters — is whether the titanium hip bracket can absorb a 15-kilonewton impulse without fracturing, whether the CFRP femur link springs back or shatters, whether the TPU outer shell distributes that energy over a large enough contact patch to protect both the machine and anything standing next to it.
Materials science is unglamorous. It does not make headlines. But in those 0.4 seconds, it is everything.
This is Article 5 in the Open Humanoid series. Articles 3 and 4 established our kinematic architecture — 39 degrees of freedom, quasi-direct-drive motors, a 1 kHz EtherCAT control loop, and a total system mass target of 80 kg. Now we ask the foundational question that all that elegant motion planning rests on: what, physically, is this robot made of?
The Mass Budget: A Constrained Life
Start with the brutal arithmetic. We have 80 kilograms to work with. From Article 4, we know that 39 QDD actuators — spanning four torque classes, from 0.4 kg fingers to 1.8 kg knee drives — consume approximately 28 kg of that budget. Batteries, sized for 60-plus minutes of normal operation at 48 V, demand another 15 kg. Compute and sensors (onboard GPU, IMU, depth cameras, force-torque units) add 4 kg.
That leaves 33 kilograms for everything structural: the bones, the joints, the skin, the wiring harnesses, the fasteners, the miscellaneous brackets that hold a complex machine together.
xychart-beta
title "Open Humanoid — 80kg Mass Budget"
x-axis ["Motors (39 QDD)", "Battery (48V)", "Structure", "Act. Housings", "Compute/Sensors", "Misc"]
y-axis "Mass (kg)" 0 --> 35
bar [28, 15, 18, 8, 4, 7]
Structural members — the limb links, spine, pelvis frame — get 18 kg. Actuator housings take 8 kg. Miscellaneous (wiring, fasteners, thermal paste, end-stops, cable management) claims 7 kg. Nothing is wasted. Every gram is assigned.
This is what engineers call a mass-constrained design problem, and it is one of the most creatively demanding in robotics. You cannot just make things heavier to make them stronger. The moment you exceed 80 kg, you ruin your gait dynamics, your motor torque margins, your fall recovery. Lightness and strength must coexist — which is precisely why material selection is not a detail. It is the central decision.
The Materials: A Cast of Characters
Carbon Fiber Reinforced Polymer — The Aerospace Choice
CFRP is what you reach for when mass is the enemy and stiffness is the goal. A well-laid carbon fiber laminate achieves a specific stiffness (Young’s modulus divided by density) of 100–150 GPa·cm³/g, compared to roughly 25 for 7075 aluminum and 27 for titanium Ti-6Al-4V. For limb links — the structural tubes that span knee to hip, elbow to shoulder — this is transformative. A CFRP tube carrying the same bending load as an equivalent aluminum section can weigh 60% less.
The catch is brittle failure. Under impact, CFRP does not yield plastically the way metals do. It delaminates. Fibers shear. The failure mode is sudden and can be catastrophic rather than graceful. This makes CFRP unsuitable for joint interfaces and impact zones — but ideal for the straight spans between joints where bending dominates and impact is less likely. The Berkeley Humanoid Lite project demonstrated this principle directly, using carbon fiber tubes as primary leg link extensions to achieve dramatic mass reduction while maintaining structural integrity under normal gait loads (arXiv:2504.17249, 2025).
CFRP also has a manufacturing challenge: you cannot machine it like metal. Every joint interface requires inserts — typically aluminum or titanium bonded and potted into the tube ends. This adds complexity and potential failure modes at every connection.
7075-T6 Aluminum — The Workhorse
Aluminum is the reason most robots do not crumble under their own weight. Specifically, 7075-T6 — the aerospace alloy with zinc as its primary alloying element — achieves a yield strength of 503 MPa at a density of 2.81 g/cm³. It machines beautifully, accepts standard anodizing for corrosion resistance, and when it fails, it fails predictably through plastic deformation rather than brittle fracture.
For actuator housings, joint brackets, and structural plates, 7075-T6 is the default choice. It is not as light as CFRP and not as strong as titanium, but it is manufacturable, reliable, and — critically — inexpensive compared to either. A kilogram of 7075 aluminum stock costs less than $10. The same titanium costs $40–80.
The distinction between 6061 and 7075 matters here. 6061-T6 (yield: 276 MPa) is the material for non-critical brackets, cover plates, and secondary structure. 7075-T6 is reserved for load-bearing geometry where the extra strength justifies the slightly higher cost and marginally reduced corrosion resistance.
Titanium Ti-6Al-4V — The Premium Option
There is something almost poetic about using Ti-6Al-4V in a humanoid robot. This alloy — 90% titanium, 6% aluminum, 4% vanadium — is also the material of choice for orthopedic implants. It is biocompatible, as it turns out, because the human body does not reject it. A humanoid robot is not a biological system, obviously, but the material properties that make it suitable for living inside a femur make it equally attractive for living as one.
Ti-6Al-4V achieves a yield strength of 880 MPa at a density of 4.43 g/cm³ — a specific strength superior to both CFRP and aluminum for ductile, impact-resistant applications. It does not delaminate. It does not yield as dramatically as aluminum. Its fatigue life under cyclic loading is exceptional, which matters enormously for a joint structure that will execute millions of steps over its operational lifetime.
The cost is real: titanium components machined from billet stock can run 5–10 times the cost of equivalent aluminum parts. We use titanium selectively — hip joint interfaces, knee load paths, and ankle brackets — the structures that see both the highest static loads and the most severe impact events during falls.
TPU and Elastomers — The Skin
A robot without soft outer surfaces is a collision hazard. ISO/TS 15066 specifies biomechanical force limits for human-robot contact in collaborative environments — for the chest, limits on quasi-static clamping force are around 210 N, and on transient impact force around 280 N. A rigid aluminum shell cannot attenuate a 15 kN impact impulse to those levels through geometry alone.
Thermoplastic polyurethane (TPU) is the structural skin of choice: an elastomer with durometer values typically ranging from Shore 40A (very soft, foam-like energy absorption) to Shore 95A (firm, impact-resistant). For the Open Humanoid outer shell, we target Shore 60–75A — soft enough to deform on contact, stiff enough to maintain shape during normal operation and resist abrasion over thousands of use cycles. Recent work on adaptive robot skin architectures demonstrates that compliant outer surfaces can simultaneously absorb collisions and host sensing elements (arXiv:2409.06369, 2024) — a future direction this platform will pursue.
xychart-beta
title "Material Comparison — Key Properties (Normalized 0-10)"
x-axis ["Specific Stiffness", "Impact Ductility", "Machinability", "Cost Efficiency", "Fatigue Life"]
y-axis "Score" 0 --> 10
bar [9.5, 2.0, 3.0, 4.0, 7.0]
The bar chart above reflects CFRP scores. For 7075-Al: [6.5, 7.5, 9.0, 9.5, 7.5]. For Ti-6Al-4V: [8.0, 8.5, 5.0, 2.5, 9.5]. Material selection is never a single dimension.
Where each material lives in the robot:
+-------------------------------+
| HEAD / NECK | 7075-Al housings, TPU cover
| (Al + TPU shell) |
+-------------------------------+
| SHOULDER JOINT BLOCK | 7075-Al housing, Ti inserts
+--------------+----------------+
| UPPER ARM | UPPER ARM | CFRP tubes (primary links)
| (CFRP link) | (CFRP link) | Al end-caps, Ti joint inserts
+--------------+----------------+
| TORSO / SPINE | 7075-Al frame, CFRP panels
| (structural core) |
+--------------+----------------+
| HIP (Ti) | HIP (Ti) | Ti-6Al-4V brackets (highest load)
+--------------+----------------+
| THIGH (CFRP)| THIGH (CFRP) | CFRP primary, Al end fittings
+--------------+----------------+
| KNEE (Ti) | KNEE (Ti) | Ti-6Al-4V, sealed bearings
+--------------+----------------+
| SHIN (CFRP) | SHIN (CFRP) | CFRP primary, Al end fittings
+--------------+----------------+
| ANKLE (Ti) | ANKLE (Ti) | Ti-6Al-4V, IP54 sealed
+--------------+----------------+
| FOOT (Al + TPU sole) | Al structure, TPU impact pad
+-------------------------------+
Outer shell: TPU 60-75A Shore throughoutStress Analysis: The Numbers Behind the Decisions
Static Loading — Standing Still
At rest, the robot’s full 80 kg bears down through the leg structure. With both feet planted, each leg carries approximately 400 N. The critical cross-section is the thigh link — a CFRP tube of roughly 50 mm outer diameter and 3 mm wall thickness, loaded primarily in bending during stance phase deviations.
The bending stress at the thigh link mid-span, assuming a 400 N axial load with a 20 mm lateral eccentricity:
sigma = M * c / I
Where:
- M = 400 N x 0.020 m = 8 N*m
- c = outer radius = 0.025 m
- I = pi(r4outer – r4inner)/4 = pi(0.025^4 – 0.022^4)/4 ~ 1.47e-7 m^4
- sigma = 8 * 0.025 / 1.47e-7 ~ 1.36 MPa
A number well within CFRP’s allowable stress of ~600 MPa in the fiber direction. Static loading is almost never the governing case. The dynamic loads are another matter entirely.
Dynamic Loading — Walking
Ground reaction forces during normal walking peak at approximately 1.2 times body weight — roughly 940 N — during mid-stance, and spike to 3 times body weight at heel strike: approximately 2,350 N. This is consistent with established biomechanical data and recent computational studies on humanoid locomotion force modeling (arXiv:2602.03177, 2026).
At 2,350 N heel-strike load, the ankle Ti-6Al-4V bracket sees a bending moment. With a 30 mm moment arm and a section modulus of approximately 3,000 mm³ for a 20 x 30 mm rectangular Ti section:
sigma = F L / Z = (2350 30) / 3000 ~ 23.5 MPa
Ti-6Al-4V yield strength: 880 MPa. Safety factor at walking: approximately 37. The ankle bracket appears massively overdesigned. It is. Until you consider what happens next.
Worst Case — The Fall
When an 80 kg robot falls from standing height (center of mass at approximately 0.9 m), it arrives at the floor with:
v = sqrt(2 g h) = sqrt(2 9.81 0.9) ~ 4.2 m/s
The impact force depends critically on the stopping time. For a rigid robot hitting a hard floor with no protective geometry, stopping time might be as short as 5 ms. Using the impulse-momentum theorem:
F = m deltav / deltat = 80 4.2 / 0.005 ~ 67,200 N
Sixty-seven kilonewtons. This is why falls matter. This is why material selection matters. Our Ti ankle bracket, rated comfortably at 23.5 MPa under walking loads, would now see:
sigma_fall = 67200 * 30 / 3000 ~ 672 MPa
That is 76% of Ti-6Al-4V’s yield strength in a single event. The safety factor collapses from 37 to just 1.3. With material variability and manufacturing tolerances accounted for, uncontrolled falls from standing height represent a genuine structural failure risk — which is precisely why fall protection algorithms, compliant foot soles, and soft outer shells are not optional features. They are structural design requirements.
Our design target: 2.5x minimum safety factor on all structural members under the maximum expected operational load, with falls mitigated by TPU sole energy absorption (targeting at least 30% impact velocity reduction) and fall-protection trajectory generation that distributes contact over larger surface areas.
Joint Design: Where Bones Meet
A humanoid joint is a bearing problem wrapped in a sealing problem wrapped in a load distribution problem.
Every rotating joint in the Open Humanoid platform uses crossed-roller bearings or angular contact bearing pairs. Crossed-roller bearings are preferred for their ability to carry combined axial, radial, and moment loads in a compact package — ideal for hip and knee joints where loading is complex and space is limited. Angular contact pairs, back-to-back configured for stiffness, suit high-speed, lower-moment joints like wrists.
Preload is the invisible variable that kills robot joints. Too little preload and the bearing develops play — backlash that compounds through the kinematic chain into positioning error and structural looseness. Too much and the bearing runs hot, the seal wears prematurely, and fatigue life collapses. We specify preload in the medium range — typically 5–10% of the basic dynamic load rating — and require all bearing assemblies to be thermally characterized before installation.
IP54 sealing is a constraint, not a preference. The IP54 rating (dust-tight enough to prevent harmful ingress, splash-proof against water from any direction) requires labyrinth seals plus elastomeric lip seals at every rotating interface. For a 39-DOF robot, this means 39 seal assemblies, each adding mass, friction, and manufacturing complexity. Nitrile rubber (NBR) seals are the default; for ankle and foot joints exposed to higher contamination risk, fluoroelastomer (FKM/Viton) provides better chemical resistance.
The Skin Problem
Here is a fact that most robot designers discover too late: a rigid robot operating near humans is a liability, not a product. ISO/TS 15066 defines biomechanical limits for collaborative robot contact that are surprisingly strict — the chest has a quasi-static clamping force limit of 210 N and a transient force limit of 280 N. These limits exist because humans are soft and robots, historically, have not been.
The solution is not padding for aesthetics. It is engineered compliance. A TPU outer shell at Shore 65A, 8 mm thickness, covering major body segments, can reduce peak contact force by 40–60% during a transient collision — not by absorbing energy passively, but by extending the collision time, which reduces peak force for the same impulse. This is the same physics that governs our fall calculation above. Recent work formalizing this relationship in the context of evolving ISO standards appears in arXiv:2602.17822 (Hartmann et al., 2026), a comparative analysis of ISO 10218 and ISO/TS 15066 in the context of next-generation collaborative robots.
Durometer selection is a genuine trade-off. Softer TPU (Shore 40A) maximizes energy absorption but tears under repetitive abrasion. The ankle and foot must therefore be harder — Shore 85–90A — because they contact the floor thousands of times per day. Hand palms need grip texture at Shore 50A. The forearm and torso shells, less likely to contact surfaces repeatedly, afford Shore 60–65A for maximum human-safety compliance.
Additive Manufacturing: Print What You Can, Machine What You Must
3D printing has changed structural prototyping in robotics irreversibly. But it has not — yet — replaced machining for primary load-bearing structure.
What we print:
- Prototype brackets and adapters in PLA or PETG for fit checking and kinematic validation
- Final-geometry non-structural parts in nylon PA12 (SLS) for cable guides, sensor mounts, and cosmetic panels
- TPU skin panels in flexible filament (FDM) or cast polyurethane for the outer shell
- Custom geometries with internal lattice structures for lightweight secondary brackets
What must be machined:
- All primary structural members with defined load paths
- Joint housings that interface with bearings (dimensional tolerance requirements of +/-0.01 mm or better)
- All titanium components (CFRP is layered, not printed into load-bearing geometry)
- Any part with fatigue life requirements under cyclic loading
The Berkeley Humanoid Lite (arXiv:2504.17249) demonstrated that 3D printing can carry more structural responsibility than traditionally assumed — their platform uses printed polymer housings for actuator integration with acceptable results under normal operation. The caveat is always impact: printed polymer, even high-density nylon, fails brittlely under concentrated impact loads in ways that machined metal does not. Our structural philosophy: print for complexity, machine for strength.
Structural Load Path
How forces actually flow through the robot body is as important as the material choices themselves. At heel strike, 2,350 N enters through the TPU sole, compresses into the aluminum foot plate, transfers through the ankle Ti bracket into the shin CFRP link, through the knee Ti bracket into the thigh CFRP link, through the hip Ti bracket into the pelvis Al frame, and up the spine to distribute across the torso structure.
graph TD
A["Foot Contact
2350 N heel strike"] --> B["TPU Sole
Energy absorption ~30%"]
B --> C["Al Foot Plate
7075-T6"]
C --> D["Ti Ankle Bracket
23.5 MPa walking / 672 MPa fall"]
D --> E["CFRP Shin Link
Axial + bending load"]
E --> F["Ti Knee Bracket
Highest moment arm"]
F --> G["CFRP Thigh Link
Primary compression member"]
G --> H["Ti Hip Bracket
Load distribution to pelvis"]
H --> I["Al Pelvis Frame
7075-T6, bilateral load merge"]
I --> J["Al/CFRP Spine
Distributed to torso mass"]
J --> K["Torso Structure
Load resolved to inertia"]
This load path drives the material choices: titanium at every joint interface (stress concentration points), CFRP for every spanning link (pure structural efficiency), aluminum at complex geometry nodes where machining enables precise interfaces.
Bill of Materials Preview: Structural Subsystem
| Component | Material | Unit Mass (kg) | Qty | Total (kg) |
|---|---|---|---|---|
| Thigh links (L+R) | CFRP tube 50mm OD | 0.35 | 2 | 0.70 |
| Shin links (L+R) | CFRP tube 40mm OD | 0.25 | 2 | 0.50 |
| Upper arm links (L+R) | CFRP tube 35mm OD | 0.18 | 2 | 0.36 |
| Forearm links (L+R) | CFRP tube 30mm OD | 0.14 | 2 | 0.28 |
| Hip brackets (L+R) | Ti-6Al-4V | 0.45 | 2 | 0.90 |
| Knee brackets (L+R) | Ti-6Al-4V | 0.38 | 2 | 0.76 |
| Ankle brackets (L+R) | Ti-6Al-4V | 0.28 | 2 | 0.56 |
| Shoulder blocks (L+R) | 7075-Al | 0.42 | 2 | 0.84 |
| Elbow blocks (L+R) | 7075-Al | 0.28 | 2 | 0.56 |
| Pelvis frame | 7075-Al | 2.20 | 1 | 2.20 |
| Torso frame | 7075-Al + CFRP panels | 3.50 | 1 | 3.50 |
| Spine links | 7075-Al | 0.60 | 1 | 0.60 |
| Head/neck structure | 7075-Al | 0.45 | 1 | 0.45 |
| Foot structures (L+R) | 7075-Al + TPU sole | 0.55 | 2 | 1.10 |
| TPU outer shell panels | TPU Shore 65A | 0.80 | 1 set | 0.80 |
| Bearings (all joints) | 52100 Steel | 0.12 | 39 | 4.68 |
| Fasteners + inserts | Ti + stainless | — | — | 0.80 |
| Wiring + thermal mgmt | Various | — | — | 0.62 |
| TOTAL | ~18.2 kg |
The 18.2 kg structural mass fits within the 18 kg target with a small margin for iteration. Actuator housings (the 8 kg line item) are budgeted separately and included in motor integration, not structural BOM.
What Comes Next
Article 6 will address the sensing subsystem — the proprioceptive, exteroceptive, and force-sensing infrastructure that gives this structural skeleton its awareness. A body without sensation is just sculpture.
The structural specification established here — CFRP for primary links, 7075-Al for complex geometry nodes, Ti-6Al-4V at high-stress joint interfaces, TPU for outer skin — is the baseline that sensing, actuation, and compute integration will now build upon. The mass budget closes. The load paths are defined. The materials are chosen.
The robot can now survive a fall. That, as it turns out, is only the beginning of the interesting engineering problems.
References
- arXiv:2504.17249 — Berkeley Humanoid Lite: Open-source, accessible, and customizable 3D-printed humanoid robot with carbon fiber structural links (2025)
- arXiv:2602.03177 — Estimation of Ground Reaction Forces from Kinematic Data during Locomotion (2026)
- arXiv:2602.17822 — Evolution of Safety Requirements in Industrial Robotics: Comparative Analysis of ISO 10218-1/2 (2011 vs. 2025) and Integration of ISO/TS 15066 (2026)
- arXiv:2409.06369 — Adaptive Electronic Skin Sensitivity for Safe Human-Robot Interaction (2024)
- arXiv:2502.01256 — Soft is Safe: Human-Robot Interaction for Soft Robots, CoboSkin (2025)
- arXiv:2601.18494 — Real-Time Prediction of Lower Limb Joint Kinematics, Kinetics, and Ground Reaction Force using Wearable Sensors (2026)
- ISO/TS 15066:2016 — Robots and robotic devices: Collaborative robots — Biomechanical limits for human-robot contact
- PMC:12829565 — Bioinspired growable humanoid robot with bone-mimetic linkages and carbon-fiber structural guides (2026)
- Scientific Reports (2025) — Opportunities, challenges and roadmap for humanoid robots in construction (doi:10.1038/s41598-025-30252-6)
- Yijin Solution (2026) — Advanced Materials for Humanoid Robots: CFRP strength-to-weight 79x steel comparison