Humanoid Robotics Design

•

Designing a humanoid robot isn't merely about replicating the human form. It's about understanding the rationale behind each joint and selectively choosing what to replicate. What's the best way to do this? First principles thinking. Here's a first principles approach to mechanical design in humanoid robotics, on the Soft Robotics Podcast: @GoingBallistic5 @MarwaEldiwiny

The typical temptation is to start with “what do humans have?” and then replicate it. But that’s a poor engineering strategy. Nature is a good starting place, but it has lots of quirks based on the specific selection pressures we faced. We don't need to copy everything. E.g wisdom teeth, tail bone, etc. In robotics, every joint should be justified on functional, not anatomical grounds.

The human body is also insanely complex. It would be madness to replicate the human hand 1:1. Our wrist, palm, and fingers consist of 27 bones, 27 joints, 34 muscles, and over 100 ligaments and tendons. Achieving 27 degrees of freedom: - Fingers: 4 each (3 for extension and flexion and one for abduction and adduction) - The thumb: 5 DOF, - Wrist: 6 DOF for rotation and translation This flexibility gives us 33 highly optimised grip types - each for different situations and objects based on the mass, size and shape, rigidity, and force and precision required. Replicating this would be cool but is it practical?

As a result - “General-purpose” is a trap. Everyone assumes humanoids have to be general purpose but building for every scenario = building for none. Warehouse humanoids won't have the same requirements as home assistants or entertainment bots. Strategic pruning leads to cheaper, more efficient, and faster-to-market products.

Good design starts with three guiding questions: • What is the robot for? • What environments will it operate in? • What degrees of freedom (DOFs) are necessary to perform those tasks? The answers dictate structure, not biology.

To apply this systematically, @GoingBallistic5 uses a joint classification method called the IN-OUT Audit. Each joint or axis is evaluated: • Indispensable: Essential for task completion. • Needed: Enhances functionality but not critical. • Optional: Offers flexibility; not necessary. • Useless: Provides no benefit for the task. • Toxic: Detrimental due to added complexity or cost. This helps filter engineering decisions through a purpose-built lens.

Example: The Neck Yaw Does a material handling robot need to turn its head left & right? Nope. It can turn its torso. Or its whole body. Neck yaw adds cost, complexity, and mass for negligible gain. Verdict? Toxic. Unless you're designing for emotional signalling or theatre.

Top-Down Audit: Joint-by-Joint Evaluation The team performs an exhaustive biomechanical breakdown using the above principles:

a. Head & Neck • Head is typically a camera platform • Neck pitch (up/down): Needed. • Neck yaw (left/right): Argued as toxic due to redundancy (turning torso or camera instead). • Examples: Agility's Digit has minimal/no head motion for safety and simplicity.

b. Arms & Shoulders • Shoulders: Indispensable, with yaw and pitch critical for reaching. • Roll: Optional but helpful for complex dexterity. • Elbows: Not indispensable for positioning, but needed for reach and manoeuvrability. • Wrist: Rotation (pronation/supination): Needed. Pitch & yaw: Debated. May not be needed if redundancy exists upstream in arm/shoulder. Wrist abduction: Often optional or useless, possibly injury-inducing.

c. Torso (Spine) Twist: Optional, helps with reach. Lean (pitch): Often redundant due to leg pitch. Tilt (roll): Possibly toxic; adds complexity without significant use. • Mentioned that some bots (e.g., Digit) have no spinal DOFs.

d. Legs (Hips, Knees, Ankles) • Hip stride (pitch): Indispensable. • Hip ab/adduction: Optional but useful for balance or squatting. • Hip rotation (yaw): Needed for turning. • Knee: Needed for crouching and efficient walking but not indispensable (you can walk stiff-legged). • Ankle: Pitch (dorsiflexion): Often needed. Roll (inversion/eversion): Helps with balance on uneven terrain—optional. Yaw: Often useless/toxic.

e. Foot Design & Gait Economy • Heavy discussion of foot pronation/supination, biomechanical subtleties, and their effects on balance, fatigue, and control. • Suggests humanoid feet are a complex ballet of micro-movements; not easily replicated with rigid mechanics.

What this approach reveals is how often redundancy exists in the body. • Neck yaw vs. torso yaw • Spine flexion vs. hip pitch • Wrist DOFs vs. shoulder/arm compensation These overlaps allow you to remove some joints without losing capability.

This isn’t about minimalism for its own sake. It’s about trimming excess to reduce complexity: • Energy draw • Mass • Mechanical failure points • Training burden (especially in reinforcement learning contexts)

Many humanoid robots in development today are beginning to reflect these principles. • Tesla Optimus removed spinal pitch. • Agility Digit avoided the head altogether (this design has been updated) • Unitree minimised wrist articulation. • Figure’s torso design changed as their understanding evolved. They're not copying humans. They're solving for their use cases.

So what’s the bigger picture here? Roboticists should shift from form-first to function-first design. That means accepting that some “human” traits don’t belong. And it means justifying every joint, every motor, every axis of motion from scratch.

Full video by @GoingBallistic5 and @MarwaEldiwiny here - youtube.com/watch?v=taFkcGsIfI0&t=1609s&ab_channel=MarwaElDiwiny